ArticlePDF Available

Intelligence-Augmented Rat Cyborgs in Maze Solving

Authors:

Abstract and Figures

Cyborg intelligence is an emerging kind of intelligence paradigm. It aims to deeply integrate machine intelligence with biological intelligence by connecting machines and living beings via neural interfaces, enhancing strength by combining the biological cognition capability with the machine computational capability. Cyborg intelligence is considered to be a new way to augment living beings with machine intelligence. In this paper, we build rat cyborgs to demonstrate how they can expedite the maze escape task with integration of machine intelligence. We compare the performance of maze solving by computer, by individual rats, and by computer-aided rats (i.e. rat cyborgs). They were asked to find their way from a constant entrance to a constant exit in fourteen diverse mazes. Performance of maze solving was measured by steps, coverage rates, and time spent. The experimental results with six rats and their intelligence-augmented rat cyborgs show that rat cyborgs have the best performance in escaping from mazes. These results provide a proof-of-principle demonstration for cyborg intelligence. In addition, our novel cyborg intelligent system (rat cyborg) has great potential in various applications, such as search and rescue in complex terrains.
Content may be subject to copyright.
RESEARCH ARTICLE
Intelligence-Augmented Rat Cyborgs in Maze
Solving
Yipeng Yu
1
, Gang Pan
1
*, Yongyue Gong
1
, Kedi Xu
2
, Nenggan Zheng
2
, Weidong Hua
1
,
Xiaoxiang Zheng
2
, Zhaohui Wu
1
1College of Computer Science and Technology, Zhejiang University, Hangzhou, Zhejiang, China, 2Qiushi
Academy for Advanced Studies, Zhejiang University, Hangzhou, Zhejiang, China
*gpan@zju.edu.cn
Abstract
Cyborg intelligence is an emerging kind of intelligence paradigm. It aims to deeply integrate
machine intelligence with biological intelligence by connecting machines and living beings
via neural interfaces, enhancing strength by combining the biological cognition capability
with the machine computational capability. Cyborg intelligence is considered to be a new
way to augment living beings with machine intelligence. In this paper, we build rat cyborgs
to demonstrate how they can expedite the maze escape task with integration of machine
intelligence. We compare the performance of maze solving by computer, by individual rats,
and by computer-aided rats (i.e. rat cyborgs). They were asked to find their way from a con-
stant entrance to a constant exit in fourteen diverse mazes. Performance of maze solving
was measured by steps, coverage rates, and time spent. The experimental results with six
rats and their intelligence-augmented rat cyborgs show that rat cyborgs have the best per-
formance in escaping from mazes. These results provide a proof-of-principle demonstration
for cyborg intelligence. In addition, our novel cyborg intelligent system (rat cyborg) has great
potential in various applications, such as search and rescue in complex terrains.
Introduction
Within the past two decades, bio-robots have been realized on different kinds of creatures,
such as cockroaches [1], moths [2], beetles [3], and rats [47]. They are expected to be superior
to traditional mechanical robots in mobility, perceptivity, adaptability, and energy consump-
tion [811]. Among them, rat robots are becoming popular for their good maneuverability. To
make a rat robot, a pair of micro electrodes are implanted into the medial forebrain bundle
(MFB) of the rats brain, and the other two pairs are implanted into the whisker barrel fields of
left and right somatosensory cortices (SI). After the rat recovers from the surgery, a wireless
micro-stimulator is mounted on the back of the rat to deliver electric stimuli into the brain
through the implanted electrodes. This allows a user, using a computer, to deliver stimulus
pulses to any of the implanted brain sites remotely. Stimulation in MFB can excite the rat robot
by increasing the level of dopamine in its brain, and stimulation in the left or right SI makes
the rat robot feel as if its whiskers are touching a barrier. Before a rat robot is used for
PLOS ONE | DOI:10.1371/journal.pone.0147754 February 9, 2016 1/18
OPEN ACCESS
Citation: Yu Y, Pan G, Gong Y, Xu K, Zheng N, Hua
W, et al. (2016) Intelligence-Augmented Rat Cyborgs
in Maze Solving. PLoS ONE 11(2): e0147754.
doi:10.1371/journal.pone.0147754
Editor: Nanyin Zhang, Penn State University,
UNITED STATES
Received: June 18, 2015
Accepted: January 7, 2016
Published: February 9, 2016
Copyright: © 2016 Yu et al. This is an open access
article distributed under the terms of the Creative
Commons Attribution License, which permits
unrestricted use, distribution, and reproduction in any
medium, provided the original author and source are
credited.
Data Availability Statement: All relevant data are
within the paper and its Supporting Information files.
Funding: This work has been supported by National
Key Basic Research Program of China
(2013CB329504) (http://www.973.gov.cn/English/
Index.aspx), Zhejiang Provincial Natural Science
Foundation of China (LR15F020001) (http://www.
zjnsf.gov.cn/), and Program for New Century
Excellent Talents in University (NCET-13-0521)
(http://www.moe.gov.cn/). The funders had no role in
study design, data collection and analysis, decision to
publish, or preparation of the manuscript.
navigation, a training process is usually needed to reinforce the desired behaviors (i.e. moving
ahead, turning left and turning right). In this process, the MFB stimulation acts as a reward as
well as a cue to move ahead, and the left and right SI stimulation act as the cues to turn left and
turn right respectively. In order to get the reward, the rat robot will learn to do the correct
behaviors corresponding to the cues. The references [4,12,13] provide more details. After suf-
ficient navigation training, the rat robot will move ahead in response to the Forward cue, turn
left in response to the Left cue, and turn right in response to the Right cue. In this paper, the rat
robot is referred to as rat cyborg.
One of the reasons that researchers take interest in developing rat cyborgs is that rats have
outstanding spatial localization abilities. They can find a way in an environment by orienting
themselves in relation to a wide variety of cues, including distal cues, typically provided by
vision, audition, and olfaction, as well as proximal cues, typically provided by tactile, kines-
thetic, and inertial systems [14]. Rats even have a well-developed magnetic compass sense for
spatial orientation [1517]. Rats do not build detailed geometrical representations of the envi-
ronment. They rely on the learnt associations between external perception and the pose belief
created from the self-motion cues. Studies of spatial orientation posit that rats use an inner
GPSin hippocampus and entorhinal cortex to create a cognitive map of the environment [18
22]. Furthermore, RatSLAM, a navigation model which is inspired by rats brain, can perform
simultaneous localization and mapping in real time on a mechanical robot [2325].
On the other hand, machines (or computers) are efficient in numerical computation, infor-
mation retrieval, statistical reasoning, and have almost unlimited storage. Different to rats,
machines have their own methods to explore and learn about the environment in spatial navi-
gation problems. They can capture many categories of information from the environment
through various sensors, such as range sensors, visual sensors, vibration sensors, acoustic sen-
sors, and location sensors. The information is saved and converted into a discrete and digital
form. Then the processed information will be synthesized to map the environment by different
paradigms [26,27]. Eventually, machines can choose various searching algorithms in path
planning, for example, flood-fill method, Dijkstras algorithm, Asearch algorithm, rapidly
exploring random tree, probabilistic roadmap. Machines run the sense process, map process,
and decision process in real time and in parallel. A typical demonstration of machines naviga-
tion abilities is the micromouse competition, in which a small rat-like mechanical robot
explores a 16×16 maze [28].
For a specific application, biological creatures and machines both have their own strengths
and weaknesses. In this paper, we ask the question: can biological intelligence be augmented
with the help of machine intelligence? To explore the question, a maze solving task was intro-
duced, in which three kinds of subjects (computer, rats and rat cyborgs) were asked to find
their ways from a predetermined starting position to a predetermined target position. It is a
challenge to embed the machines maze solving capability into their navigation. In our experi-
ments, the computer traversed the mazes based on an improved wall follower approach, six
rats traversed 14 mazes one by one all by themselves, and then traversed the 14 mazes again in
the same order with the assistance of the computer. Performance was measured by steps, cover-
age rates, and time spent, allowing for comparisons.
Materials and Methods
Subjects
Six rats (adult Sprague Dawley rats; 290 350 g) took part in the experiments over the course of
one month. All of the six rats (named DH12, SV15, SV17, MV12, M01 and M03) had already
gone through a sufficient navigation training process, which means they would turn left in
Rat Cyborgs in Maze Solving
PLOS ONE | DOI:10.1371/journal.pone.0147754 February 9, 2016 2/18
Competing Interests: The authors have declared
that no competing interests exist.
response to the Left stimulus, turn right in response to the Right stimulus, and move ahead in
response to the Forward stimulus. Note that in this study, Left and Right stimuli are used to
instruct the rat cyborgs to turn left and right, as well as to prevent them from moving into dead
roads. Forward stimulus is not used to instruct them to move ahead, but to reward them. In our
experiments, the six rats and the six rat cyborgs are the same rats. A rat solving mazes by itself is
called a rat, while the same rat solving mazes with the assistance of computer via its backpack
stimulator is called a rat cyborg. This study was approved by the Ethics Committee of Zhejiang
University (Agreement number: Zju201402-1-02-034). All of the rats used in these experiments
were well cared for by the animal keepers. All of the experiments were performed in accordance
with the guidelines issued by the Ethics Committee of Zhejiang University, and they complied
with the China Ministry of Health Guide for the Care and Use of Laboratory Animals.
Apparatus
The experimental system for maze solving is shown in Fig 1. The maze is made of wood, and
comprises 10×10 unit squares (15 cm×15 cm per unit square). Walls of this maze are also 15
cm high and the outside walls enclose the entire maze. The maze is covered with a piece of per-
spex on the top, which is used to prevent rats from climbing and escaping. The starting posi-
tion S is in the bottom-left cell of the maze, and target position G is in the top-right cell of the
maze. A notable feature of this maze is that the four walls of each cell can be inserted or
removed. Therefore we can change the layout of each maze, and set various paths from S to G.
In our experiments, we designed 14 mazes (maze 1 to maze 14, see S1 Fig) with different levels
of complexity. Furthermore, there were a box of peanut butter and a dish of water in the target
cell. The peanut butter was used as an odor source, and the water was used as a reward.
The whole experimental process was captured by a web camera. The camera we used could
capture video clips at a rate of 15 frames per second, with a resolution of 640 by 480 pixels. The
computer acquired the explored maze information and motion states of the rats in real time
with the assistance of the camera. The computer-aided maze solving system running on the
computer was implemented in C++ with a user-friendly interface. Left, Right and Forward
Fig 1. Experimental system for maze solving.
doi:10.1371/journal.pone.0147754.g001
Rat Cyborgs in Maze Solving
PLOS ONE | DOI:10.1371/journal.pone.0147754 February 9, 2016 3/18
stimuli were sent to the backpack of a rat via Bluetooth. Left and Right stimuli were sent to
guide the rats only in the maze solving procedure by rat cyborgs. Forward stimulus acted as
another reward (like water) after the rats reached the target cell in maze solving procedures by
rats and by rat cyborgs.
Procedures
Maze Solving Training. In this procedure, in order to get the rewards in the target cell,
rats had to learn to find a path from the constant starting position to the constant target posi-
tion. We used double rewards to reinforce rats maze solving behavior: water reward and elec-
tric reward of MFB stimulation. Water is a natural reward, which is indispensable to rats. MFB
reward is a train of short-duration pulses, which can excite the rats. A combination of water
reward and MFB reward can accelerate this maze solving training procedure [29,30]. In addi-
tion, olfactory cues can be an aid for the rats in solving spatial problems [3134]. In our experi-
ments, a box of peanut butter was placed in the target cell (see Fig 1), which was used to help
rats locate the target cell, and activate their odor tracking abilities.
The detailed training procedure is as follows. After water deprivation for 24 hours, six rats,
which had been well trained in the navigation training, proceeded to the maze solving training,
which consisted of 3 to 5 random mazes per day in consecutive days. The training time in each
day was 8-12 a.m. and 2-6 p.m. In each maze, at first the rat was placed in the constant starting
cell. It had to learn to find a path to the target cell by itself. After the rat arrived at the target cell
and drank a drop of water (0.15 ml) in a dish, five consecutive MFB rewards were sent. Each
rat ran each maze only one time, and each maze had a different layout. Once all of the six rats
finished a maze, they proceeded to training in the next maze. Each rat was offered only 6 ml
water after one days training. This amount of water was to meet the basic need, but not enough
to completely satisfy a rat. The rat still wanted to drink water on the next day. The entire train-
ing lasted 7 days. After this procedure, all of the six rats had a strong desire to run in the mazes
to search for the water and MFB rewards.
Maze Solving by Computer. Similar to rats, the computer had to find a path from the
start to the target with no knowledge of the mazes. There are a number of different maze solv-
ing algorithms, such as random mouse, wall follower, Pledge, and Tremauxs algorithm [35,
36]. In this study, on the basis of dead road detection and unique road detection, two wall fol-
lower rules (left-hand and right-hand) were adopted to traverse the 14 mazes. Dead roads refer
to roads that can not lead to the target. There are two kinds of dead roads: visited dead road
and non-visited dead road. The dead road detection algorithm is listed in Algorithm A in S1
File. A unique road refers to the only way to the target. The detailed unique road detection
algorithm is listed in Algorithm B in S1 File. The complete maze solving algorithm is listed in
Algorithm 1. For left-hand rule, the traversal sequence is clockwise (left!front!right!back),
and for right-hand rule, the traversal sequence is anticlockwise (right!front!left!back). Fig
2shows two maze solving processes by our algorithm. Steps and coverage rates of maze solving
in the 14 mazes were recorded. We took the average steps and average coverage rates of the
left-hand rule and right-hand rule as the performance measures. Two demo videos of maze
solving by our algorithm are presented (Videos A and B in S2 File).
Algorithm 1: Maze solving by our algorithm (left-hand).
1while the current cell C is not the target cell do
2if there is a unique cell U which is accessible in the four adjacent cells
then
3 move to U;
4end
Rat Cyborgs in Maze Solving
PLOS ONE | DOI:10.1371/journal.pone.0147754 February 9, 2016 4/18
5else if the left cell of C is accessible and not a dead cell then
6 move to the left cell;
7end
8else if the front cell of C is accessible and not a dead cell then
9 move to the front cell;
10 end
11 else if the right cell of C is accessible and not a dead cell then
12 move to the right cell;
13 end
14 else if the back cell of C is accessible and not a dead cell then
15 move to the back cell;
16 end
17 end
To ensure that our algorithm can reflect the computers maze solving capacity, we compare
its performance in 24 mazes (maze 1 to maze 24, see S1 Fig) with that of three classical maze
solving algorithms (i.e. wall follower, Pledge and Tremauxs algorithm). The experimental
results are presented in Fig 3. We can see from the left column that both the steps and coverage
rates of our algorithm in most mazes are less than those of the other three algorithms. From
the right column, the average steps and average coverage rates of our algorithm are less than
those of the other three algorithms. The statistical analysis (two-tailed paired t-test) also shows
that there are significant differences of the steps and coverage rates between our algorithm and
the other three algorithms (p<0.001). The results indicate that our algorithm outperforms the
other three algorithms in maze solving. Thus we take the performance of our algorithm as
computers performance in maze solving, and compare it with that of rats and rat cyborgs.
Maze Solving by Rats. After the maze solving training, the entire maze was washed and
dried, and then the six rats proceeded to this formal maze solving procedure. In this procedure,
the rats solved 14 mazes one by one using their own spatial learning abilities. Details of this
procedure are similar to the maze solving training procedure. The experimental time in each
day was 8-12 a.m. and 2-6 p.m. Each rat ran each maze only one time. In each maze, at first the
rat was placed in the constant starting cell. After the rat arrived at the target cell and drank a
Fig 2. Maze solving by our algorithm. The yellow cell is the current position of the explorer. Blue cells
indicate a unique road to the target. Cyan cells are cells which have not been explored, and walls of these
cells now are unknown to the explorer. Pink slash (n) denotes the non-visited dead road, red cross (×)
denotes the visited dead road. (a) Left-hand. (b)Right-hand.
doi:10.1371/journal.pone.0147754.g002
Rat Cyborgs in Maze Solving
PLOS ONE | DOI:10.1371/journal.pone.0147754 February 9, 2016 5/18
drop of water (0.15 ml) in a dish, five consecutive MFB rewards were sent. Once all of the six
rats finished a maze, they proceeded to solve the next maze. Each rat was offered 6 ml water at
the end of one days experiment. Steps, coverage rates and time spent of maze solving by the
rats in the 14 mazes were analysed. Additionally, we observed that the strategy rats took to
solve the mazes was not easily understood (in other words, it was not perfect). They revisited
the visited cells, explored the dead roads which could be easily detected by the computer, and
even returned to the starting cell when they were confused. This offered an opportunity to the
computer to help rats in maze solving. A demo video of this procedure is shown in Video C in
S2 File.
Maze Solving by Rat Cyborgs. A rat cyborg is a computer-aided rat, which is a combina-
tion of a rat and machines. After the procedure of maze solving by rats, the entire maze was
washed and dried, and then the same six rats traversed the 14 mazes in the same order with the
computers help. Details of this procedure are similar to the procedure of maze solving by rats,
except that the rats have the help of the computer to solve the mazes. Fig 4 shows two maze
solving processes by rat cyborgs. With the motivation to help rats solve the mazes, the com-
puter tracked the rats, analyzed the explored maze information, and decided when and how to
intervene. The computer aided the rats under three rules: (1) if there was a path to the unique
road, the computer would find the shortest path, then Left and Right commands would be sent
to navigate the rat to the unique road; (2) if the rat was going to enter a dead cell, Left or Right
Fig 3. Steps and coverage rates of maze solving by our algorithm, wall follower, Pledge and Tremauxs algorithm. Results are presented as mean±s.
e.m. on the right. *p<0.05, **p<0.01, ***p<0.001.
doi:10.1371/journal.pone.0147754.g003
Rat Cyborgs in Maze Solving
PLOS ONE | DOI:10.1371/journal.pone.0147754 February 9, 2016 6/18
commands would be sent to prevent such a move; (3) if the rat was in a loop (detected by Algo-
rithm C in S1 File), the computer would find the shortest path to the current destination, then
Left and Right commands would be sent to navigate the rat to follow the path. In this experi-
ment, the computer automatically analyzed the dead roads, unique roads, loop roads and
shortest roads in real time, and generated the guiding information to overlay on the live video,
which is playing on the screen of the computer. At the same time, an expert was watching the
guiding information shown on the screen and sent the Left and Right commands to direct the
rat under the three rules. The sent Left and Right commands were also shown simultaneously
in the video. Please note that the expert was not directly involved in the maze solving, his/her
operations to send the Left and Right commands under the guidance of the computer is just a
simple replacement of the computers execution (action), not an augmentation or a diminution
of the intelligence, since all the path detection are performed by the computer. Apart from the
above three rules, rats solved the mazes of their own free will. We did not treat rats as machines
which were totally controlled by the computer. Steps, coverage rates and time spent of maze
solving by the rat cyborgs in the 14 mazes were analysed. A demo video of this procedure is
shown in Video D in S2 File.
Algorithm 2: Rat cyborg tracking algorithm.
1 Save a background image to I
a
;
2while the rat cyborg has not reached the target cell G do
3 pick an image I
b
from the camera;
4 use cvAbsDiff to subtract I
a
from I
b
, write the difference to I
c
;
5 use cvThreshold to get a binary image I
d
from I
c
;
6 use cvFindContours to find all of the contours C
o
in I
d
;
7 search the largest contour C
max
in C
o
;
8 calculate the centroid of pixels in C
max
, write it to the body position R
b
;
Fig 4. Maze solving by rat cyborgs. The red point is the body position, the black point is the head position,
and the blue line is the heading direction of the rat cyborg. Our rat cyborg tracking method (see Algorithm 2) is
implemented by OpenCV. Blue cells indicate a unique road to the target. Red cells are cells which have not
been explored, and walls of these cells now are unknown to the rat cyborg. The pink slash (n) denotes the
non-visited dead road, the red cross (×) denotes the visited dead road, and the yellow gdenotes entrances
to the unexplored area. Consecutive blue points indicate the shortest path to the current destination. The
shortest path is detected by A*algorithm [37].
doi:10.1371/journal.pone.0147754.g004
Rat Cyborgs in Maze Solving
PLOS ONE | DOI:10.1371/journal.pone.0147754 February 9, 2016 7/18
9 use cvGoodFeaturesToTrack to detect corners in C
max
find a 40×40 square
S
o
in C
max
;
10 find a 40×40 square S
o
in C
max
, which has the most corners;
11 calculate the centroid of corners in S
o
, write it to the head position R
h
;
12 end
Results
In this section, the experimental results of maze solving by computer, by rats and by rat cyborgs
are presented. We first compare steps and coverage rates of maze solving among them. Then
we compare time spent of maze solving between rats and rat cyborgs. Finally, correlation
between the steps and correlation between the coverage rates are analyzed.
Steps
Steps are the sum of times visiting each cell. A cell can be visited multiple times. Steps of maze
solving by the six rats/rat cyborgs are presented with those by the computer in Fig 5. We can
see from the left column that the steps of rat cyborgs in most mazes are less than those of rats,
and may be equivalent to those of the computer. From the right column, the average steps of
each rat cyborg are less than that of each corresponding rat: DH12 (from 117.71±25.70 to
78.64±14.94), SV15 (from 92.57±15.28 to 64.43±9.12), SV17 (from 80.43±12.92 to 67.36
±8.41), MV12 (from 112.86±8.74 to 82.86±4.96), M01 (from 134.00±21.53 to 69.29±9.88), and
M03 (from 156.43±33.92 to 64.57±10.06); and the average steps of the computer (77.79±4.90)
are also less than that of each rat. The two-tailed paired t-test shows that between computer
and rats, there are three rats (i.e. MV12, M01 and M03) that performed worse than the com-
puter; between computer and rat cyborgs, there are no statistically significant differences; and
between rats and rat cyborgs, there are five rats (i.e. DH12, SV15, MV12, M01 and M03) that
performed better with the assistance of the computer. These results suggest that in maze solv-
ing, based on the steps, the performance of rat cyborgs is better than that of individual rats, and
comparable to that of individual computer.
Coverage Rates
Coverage rate is the number of visited cells. Every visited cell is counted only one time, so the
maximum value of the coverage rate is 100. The larger the coverage rate is, the less efficient the
maze solving is. It can evaluate the performance of the explorer in maze solving from another
point of view. Coverage rates of maze solving by the six rats/rat cyborgs are presented with
those by the computer in Fig 6. We can see from the left column that the coverage rates of rat
cyborgs in most mazes are less than those of the computer and those of the rats. From the right
column, the average coverage rates of each rat cyborg are less than that of each corresponding
rat: DH12 (from 58.71±6.59 to 52.21±6.89), SV15 (from 55.36±6.26 to 51.71±6.84), SV17
(from 53.57±7.06 to 51.64±6.35), MV12 (from 60.36±5.25 to 54.57±5.20), M01 (from 65.07
±5.86 to 52.71±6.92), and M03 (from 64.86±7.27 to 51.07±6.92); and the average coverage
rates of each rat cyborg are also less than that of the computer (60.50±5.13). The two-tailed
paired t-test shows that between computer and rats, there are no statistically significant differ-
ences; between computer and rat cyborgs, all of the six rat cyborgs performed better than the
computer; and between rats and rat cyborgs, there are four rats (i.e. DH12, MV12, M01 and
M03) that performed better with the assistance of the computer. These results suggest that in
maze solving, based on the coverage rates, the performance of rat cyborgs is better than that of
individual computer and that of individual rats.
Rat Cyborgs in Maze Solving
PLOS ONE | DOI:10.1371/journal.pone.0147754 February 9, 2016 8/18
Time Spent
Time spent measures the period between a rat/rat cyborg starts to move at S and the rat/rat
cyborg reaches G. Besides steps and coverage rates, we further compare the time spent of the
rats in the 14 mazes with those of the rat cyborgs. Time spent of maze solving by the six rats/
rat cyborgs are presented in Fig 7. As we can see from the left column that the time spent of rat
cyborgs in most mazes are less than those of rats. From the right column, the average time
spent of each rat cyborg is less than that of each corresponding rat: DH12 (from 168.07±44.95
to 129.36±32.74), SV15 (from 347.43±68.68 to 125.00±31.48), SV17 (from 328.50±40.19 to
253.43±31.69), MV12 (from 588.43±106.85 to 333.07±61.58), M01 (from 156.14±33.13 to
Fig 5. Steps of maze solving by computer, rats and rat cyborgs. The steps of maze solving by DH12, SV15, SV17, MV12, M01 and M03 are shown from
the top to the bottom, respectively. Data are presented as mean±s.e.m. on the right. *p<0.05, **p<0.01, ***p<0.001.
doi:10.1371/journal.pone.0147754.g005
Rat Cyborgs in Maze Solving
PLOS ONE | DOI:10.1371/journal.pone.0147754 February 9, 2016 9/18
70.50±8.82), and M03 (from 144.36±36.35 to 45.14±13.96). The two-tailed paired t-test shows
that there are four rats (i.e. SV15, MV12, M01 and M03) that performed better with the assis-
tance of the computer. These results suggest that in maze solving, based on the time spent, the
performance of rat cyborgs is better than that of individual rats.
Fig 6. Coverage rates of maze solving by computer, rats and rat cyborgs. The coverage rates of maze solving by DH12, SV15, SV17, MV12, M01 and
M03 are shown from the top to the bottom, respectively. Data are presented as mean±s.e.m. on the right. *p<0.05, **p<0.01, ***p<0.001.
doi:10.1371/journal.pone.0147754.g006
Rat Cyborgs in Maze Solving
PLOS ONE | DOI:10.1371/journal.pone.0147754 February 9, 2016 10 / 18
Correlation
The 14 mazes have different levels of complexity from the human perspective, and the more
complicated mazes should consume more steps to solve. Maze 11 was supposed to be the most
complicated maze among the 14 mazes, and it took the rats exceptionally many more steps (see
Fig 5). Firstly, we analyze the Pearsons correlation coefficients of steps and coverage rates
Fig 7. Time spent of maze solving by rats and rat cyborgs. The time spent of maze solving by DH12, SV15, SV17, MV12, M01 and M03 are shown from
the top to the bottom, respectively. Data are presented as mean±s.e.m. on the right. *p<0.05, **p<0.01, ***p<0.001.
doi:10.1371/journal.pone.0147754.g007
Rat Cyborgs in Maze Solving
PLOS ONE | DOI:10.1371/journal.pone.0147754 February 9, 2016 11 / 18
between each other within the six rats/rat cyborgs (see Fig 8). In terms of steps, except MV12,
the other five rats/rat cyborgs have a positive correlation between each other in maze solving.
In terms of coverage rates, all of the six rats/rat cyborgs have a positive correlation between
each other in maze solving. These results indicate that a maze that is difficult for a rat/rat
Fig 8. The Pearsons correlation coefficients of steps and coverage rates between each other within the six rats/rat cyborgs. (a) Correlation
coefficients of steps within the six rats. (b) Correlation coefficients of steps within the six rat cyborgs. (c) Correlation coefficients of coverage rates within the
six rats. (d) Correlation coefficients of coverage rates within the six rat cyborgs.
doi:10.1371/journal.pone.0147754.g008
Rat Cyborgs in Maze Solving
PLOS ONE | DOI:10.1371/journal.pone.0147754 February 9, 2016 12 / 18
cyborg should be difficult for other rats/rat cyborgs as well. Secondly, we further analyze the
Pearsons correlation coefficients of steps and coverage rates between computer and rats, com-
puter and rat cyborgs, and rats and the corresponding rat cyborgs (see Table 1). In terms of
steps, all of the correlation coefficients are strong positive, except those between MV12 and the
computer. In terms of coverage rates, all of the correlation coefficients are strong positive.
These results indicate that a maze which is difficult for a rat should be difficult for the corre-
sponding rat cyborg and computer as well, and vice versa.
Discussion
The experimental results show that, in terms of steps, coverage rates and time spent, rat
cyborgs performed better than rats in maze solving. However, in terms of steps, performance
of rat cyborgs did not show remarkable advantages over that of computer. Actually, rats did
not strive to reach the destination in the least number of steps. They sometimes would revisit
visited cells again and again, and wander between adjacent cells. Nonetheless, all of the six rat
cyborgs outperformed the computer in terms of the coverage rates. Moreover, the rat cyborgs
are agile in different types of terrain (see S1 Video), and have the potential to solve unantici-
pated problems relying on instinct [38].
Because the six rats and the six rat cyborgs were the same rats, the possible interference
should be carefully prevented. In our experiments, we took four measures to avoid interference:
(1) for a rat and its rat cyborg, the layout of each maze was the same, while the physical walls of
each maze cell were changed; (2) a maze solved by a rat would be solved by its rat cyborg at
least 4 days later, not in the following day; (3) in the interval, the rat would be asked to carry
out experiments with other mazes to further weaken its memory of the previous maze; (4)
between the procedure of maze solving by rats and the procedure of maze solving by rat
cyborgs, the entire maze was washed and dried to remove the possible odor interference. In
order to verify that the performance enhancement of the rat cyborgs in maze solving was not
attributed to what the rats had experienced, we conducted two supplemental experiments.
In supplemental experiment 1, two rats (i.e. M01 and M03) first traversed 5 mazes (maze 15
to maze 19, see S1 Fig) one by one with the assistance of the computer, and then traversed the 5
mazes again in the same order all by themselves. If a strong memory of the mazes has been
gained by the rat cyborgs, it will then benefit the performance of the rats. The experimental
results are shown in Fig 9. As we can see, the steps of each rat cyborg are less than that of the
corresponding rat in each maze. The coverage rates of each rat cyborg are less than that of the
corresponding rat except in maze 17 of M03, the coverage rates of each rat cyborg are less than
that of the computer except in maze 18, and the average coverage rates of each rat cyborg
(M01: 49.40±7.72, M03: 49.20±7.51) are less than that of the corresponding rat(M01: 59.80
±9.68, M03: 56.60±8.35) and the computer (60.00±2.57). Besides, M01 spent less time in maze
Table 1. The Pearsons correlation coefficients of steps and coverage rates between computer and rats, computer and rat cyborgs, and rats and
rat cyborgs.
DH12 SV15 SV17 MV12 M01 M03
scscsc s cscsc
Computer vs Rats 0.5610 0.8822 0.6524 0.9141 0.4984 0.8951 0.1137 0.7724 0.5784 0.7966 0.6979 0.7970
Computer vs Rat cyborgs 0.5696 0.9018 0.6653 0.9022 0.6034 0.9059 -0.2160 0.8667 0.7198 0.9399 0.6548 0.9376
Rats vs Rat cyborgs 0.9254 0.9224 0.5469 0.9450 0.7024 0.9005 0.7004 0.9123 0.6302 0.7881 0.5639 0.7350
s is an abbreviation for steps, c is an abbreviation for coverage rates.
doi:10.1371/journal.pone.0147754.t001
Rat Cyborgs in Maze Solving
PLOS ONE | DOI:10.1371/journal.pone.0147754 February 9, 2016 13 / 18
solving with the assistance of the computer except in maze 17. These results show that rat
cyborgs have a better performance than rats. This is consistent with the conclusion of the previ-
ous experiments, and demonstrates that the performance enhancement of the rat cyborgs in
maze solving should not be attributed to what the rats had experienced.
In supplemental experiment 2, two rats (i.e. M01 and M03) first traversed other 5 mazes
(maze 20 to maze 24) one by one all by themselves. Then these 5 mazes were flipped along the
diagonal axis from the entrance to the exit, and traversed by the two rats in the same order
with the assistance of the computer. Note that each pair of the flipped maze and the original
one (see S1 Fig) have the same complexity but different space layouts. In this way, a rats mem-
ory of the original maze can hardly help the counterpart rat cyborg solve the flipped maze. The
experimental results are shown in Fig 10. As we can see, the steps of each rat cyborg are less
than that of the corresponding rat in each maze. The coverage rates of each rat cyborg are less
than that of the computer except in maze 24 of M01, and the average coverage rates of each rat
cyborg (M01: 46.00±8.92, M03: 36.40±3.36) are less than that of the corresponding rat(M01:
52.40±7.52, M03: 40.80±7.22) and the computer (56.60±2.96). Besides, M03 spent less time in
maze solving with the assistance of the computer except in maze 21. These results show that rat
cyborgs have a better performance than rats. This is consistent with the conclusion of the
Fig 9. Steps, coverage rates and time spent of maze solving bycomputer, M01 and M03 in maze 15 to maze 19. (a) Steps of maze solving by computer
and M01. (b) Steps of maze solving by computer and M03. (c) Coverage rates of maze solving by computer and M01. (d) Coverage rates of maze solving by
computer and M03. (e) Time spent of maze solving by computer and M01. (f) Time spent of maze solving by computer and M03.
doi:10.1371/journal.pone.0147754.g009
Rat Cyborgs in Maze Solving
PLOS ONE | DOI:10.1371/journal.pone.0147754 February 9, 2016 14 / 18
previous experiments, and also demonstrates that the performance enhancement of the rat
cyborgs in maze solving should not be attributed to what the rats had experienced.
Thanks to the rapid advance of brain-machine interfaces (BMIs), the connection and inter-
action between the organic components and computing components of the cyborg intelligent
systems are becoming deeper and better [3943]. Based on such kinds of symbiotic bio-
machine systems, a new type of intelligence, which we refer to as cyborg intelligence, will play
an increasingly crucial role. Cyborg intelligence is a convergence of machine and biological
intelligence, which is capable of integrating the two heterogeneous intelligences at multiple lev-
els [44,45]. It has the potential to deliver tremendous benefits to society, such as in search and
rescue, health care, and entertainment. In this work, the mobility, perceptibility and cognition
capability of the rats were combined with the sensing and computing power of the machines in
the rat cyborg system. The experimental results of the rat cyborg system in maze solving pro-
vide a proof-of-principle demonstration for cyborg intelligence.
Conclusions and Future Work
In this paper, we build intelligence-augmented rat cyborgs and present a comparative study of
maze solving by computer, by rats, and by rat cyborgs. Computer aids rats in dead road detec-
tion, unique road detection, loop detection, and shortest path detection. In terms of steps, cov-
erage rates and time spent, the rat cyborgs have a better performance than the individual rats
in maze solving; in terms of coverage rates, the rat cyborgs have a better performance than the
Fig 10. Steps, coverage rates and time spent of maze solving bycomputer, M01 and M03 in maze 20 to maze 24. (a) Steps of maze solving by
computer and M01. (b) Steps of maze solving by computer and M03. (c) Coverage rates of maze solving by computer and M01. (d) Coverage rates of maze
solving by computer and M03. (e) Time spent of maze solving by computer and M01. (f) Time spent of maze solving by computer and M03.
doi:10.1371/journal.pone.0147754.g010
Rat Cyborgs in Maze Solving
PLOS ONE | DOI:10.1371/journal.pone.0147754 February 9, 2016 15 / 18
individual computer in maze solving. From the systematic perspective, the rats capability of
maze solving has been augmented by the computer.
In future work, more tasks will be introduced, and the complexity of tasks will be quantified.
To avoid excessive intervention with the rats, the strength of the computers assistance will be
graded. In addition, more practical rat cyborgs will be investigated: the web camera will be
replaced by sensors mounted on rats, such as tiny camera, ultrasonic sensors, infrared sensors,
electric compass, and so on, to perceive the real unknown environment in real time; and the
computer-aided algorithms can be housed on a wireless backpack stimulator instead of in the
computer.
Supporting Information
S1 Fig. Maps of the 24 mazes solved in our experiments. The start is in the bottom-left cell of
each maze, and the target is in the top right-cell of each maze.
(PDF)
S1 File. Dead road detection (Algorithm A). Unique road detection (Algorithm B). Loop
detection (Algorithm C).
(ZIP)
S2 File. Maze solving by computer, rats and rat cyborgs. Computer solved maze 5 by our
algorithm under the left-hand rule (Video A). Computer solved maze 5 by our algorithm
under the right-hand rule (Video B). DH12 solved maze 5 by itself (Video C). DH12 solved
maze 5 with the assistance of the computer (Video D).
(ZIP)
S1 Video. A rat cyborg traversing complex terrain.
(MP4)
S1 Table. Experimental data of maze solving in the 24 mazes (maze 1 to maze 24) by our
algorithm, wall follower, Pledge and Tremauxs algorithm.
(XLS)
S2 Table. Experimental data of maze solving in the 14 mazes (maze 1 to maze 14) by the
computer and 6 rats/rat cyborgs.
(XLS)
S3 Table. Experimental data of maze solving in the 10 mazes (maze 15 to maze 24) by the
computer and 2 rats/rat cyborgs.
(XLS)
S4 Table. Stimulus parameters of the 6 rats/rat cyborgs.
(XLS)
Acknowledgments
This work is supported by National Key Basic Research Program of China (2013CB329504),
Zhejiang Provincial Natural Science Foundation of China (LR15F020001), and Program for
New Century Excellent Talents in University (NCET-13-0521). The authors are grateful to the
editor and reviewers for their insightful comments. They also are pleased to thank Liqiang
Gao, Liujing Zhuang, Shengzhang Lai for technical support and Chaonan Yu for the rat
surgery.
Rat Cyborgs in Maze Solving
PLOS ONE | DOI:10.1371/journal.pone.0147754 February 9, 2016 16 / 18
Author Contributions
Conceived and designed the experiments: YY GP KX ZW. Performed the experiments: YY
WH YG. Analyzed the data: YY GP YG. Contributed reagents/materials/analysis tools: YY GP
KX NZ XZ. Wrote the paper: YY GP.
References
1. Holzer R, Shimoyama I. Locomotion control of a bio-robotic system via electric stimulation. In: IEEE/
RSJ International Conference on Intelligent Robots and Systems, 1997. vol. 3. pp. 15141519.
2. Bozkurt A, Gilmour RF, Lal A. Balloon-assisted flight of radio-controlled insect biobots. IEEE Transactions
on Biomedical Engineering. 2009; 56(9):23042307. doi: 10.1109/TBME.2009.2022551 PMID: 19692306
3. Sato H, Berry CW, Peeri Y, Baghoomian E, Casey BE, Lavella G, et al. Remote radio control of insect
flight. Frontiers in Integrative Neuroscience. 2009; 3(2):14.
4. Talwar SK, Xu S, Hawley ES, Weiss SA, Moxon KA, Chapin JK. Rat navigation guided by remote con-
trol. Nature. 2002; 417(6884):3738. doi: 10.1038/417037a PMID: 11986657
5. Wang Y, Lu M, Wu Z, Tian L, Xu K, Zheng X, et al. Visual cue-guided rat cyborg for automatic naviga-
tion. IEEE Computational Intelligence. 2015; 10(2):4252. doi: 10.1109/MCI.2015.2405318
6. Sun C, Zheng N, Zhang X, Chen W, Zheng X. Automatic navigation for rat-robots with modeling of the
human guidance. Journal of Bionic Engineering. 2013; 10(1):4656.
7. Thomson EE, Carra R, Nicolelis MA. Perceiving invisible light through a somatosensory cortical pros-
thesis. Nature Communications. 2013; 4(2):6678.
8. Etienne S, Guthrie M, Goillandeau M, Nguyen TH, Orignac H, Gross C, et al. Easy rider: monkeys learn
to drive a wheelchair to navigate through a complex maze. PloS One. 2014; 9(5):e96275. doi: 10.1371/
journal.pone.0096275 PMID: 24831130
9. Zhang S, Mizutani A, Srinivasan MV. Maze navigation by honeybees: learning path regularity. Learning
& Memory. 2000; 7(6):363374. doi: 10.1101/lm.32900
10. Zhang S, Si A, Pahl M. Visually guided decision making in foraging honeybees. Frontiers in Neurosci-
ence. 2012; 6(6):8888. doi: 10.3389/fnins.2012.00088 PMID: 22719721
11. Xu D, Hao X, Wang Z, Duan Y, Marsh R, Yu S, et al. A virtual radial arm maze for the study of multiple
memory systems in a functional magnetic resonance imaging environment. International Journal of Vir-
tual Reality (IJVR). 2012; 11(2):6376.
12. Feng Z, Chen W, Ye X, Zhang S, Zheng X, Wang P, et al. A remote control training system for rat navi-
gation in complicated environment. Journal of Zhejiang University Science A. 2007; 8(2):323330. doi:
10.1631/jzus.2007.A0323
13. Lee MG, Jun G, Choi HS, Jang HS, Bae YC, Suk K, et al. Operant conditioning of rat navigation using
electrical stimulation for directional cues and rewards. Behavioural Processes. 2010; 84(3):715720.
PMID: 20417259
14. Sutherland RJ, Whishaw IQ, Kolb B. A behavioural analysis of spatial localization following electrolytic,
kainate- or colchicine-induced damage to the hippocampal formation in the rat. Behavioural Brain
Research. 1983; 7(2):133153. doi: 10.1016/0166-4328(83)90188-2 PMID: 6830648
15. Mather JG, Baker RR. Magnetic sense of direction in woodmice for route-based navigation. Nature.
1981; 291(5811):152155. doi: 10.1038/291152a0
16. Muheim R, Edgar NM, Sloan KA, Phillips JB. Magnetic compass orientation in C57BL/6J mice. Learn-
ing & Behavior. 2006; 34(4):366373. doi: 10.3758/BF03193201
17. Phillips JB, Youmans PW, Muheim R, Sloan KA, Landler L, Painter MS, et al. Rapid learning of mag-
netic compass direction by C57BL/6 mice in a 4-armed pluswater maze. PloS One. 2013; 8(8):
e73112. doi: 10.1371/journal.pone.0073112 PMID: 24023673
18. OKeefe J. Place units in the hippocampus of the freely moving rat. Experimental Neurology. 1976; 51
(1):78109. PMID: 1261644
19. Fyhn M, Molden S, Witter MP, Moser EI, Moser MB. Spatial representation in the entorhinal cortex. Sci-
ence. 2004; 305(5688):12581264. doi: 10.1126/science.1099901 PMID: 15333832
20. Sargolini F, Fyhn M, Hafting T, McNaughton BL, Witter MP, Moser MB, et al. Conjunctive representa-
tion of position, direction, and velocity in entorhinal cortex. Science. 2006; 312(5774):758762. doi: 10.
1126/science.1125572 PMID: 16675704
21. Crenshaw DG, Asin K, Gottschalk WK, Liang Z, Zhang N, Roses AD. Effects of low doses of pioglita-
zone on resting-state functional connectivity in conscious rat brain. PloS One. 2015; 10(2):e0117973.
doi: 10.1371/journal.pone.0117973 PMID: 25671601
Rat Cyborgs in Maze Solving
PLOS ONE | DOI:10.1371/journal.pone.0147754 February 9, 2016 17 / 18
22. Zhang N, Rane P, Huang W, Liang Z, Kennedy D, Frazier JA, et al. Mapping resting-state brain net-
works in conscious animals. Journal of Neuroscience Methods. 2010; 189(2):186196. PMID:
20382183
23. Milford MJ, Wyeth GF. Mapping a suburb with a single camera using a biologically inspired SLAM sys-
tem. IEEE Transactions on Robotics. 2008; 24(5):10381053. doi: 10.1109/TRO.2008.2004520
24. Vasile C, Pavel A, Buiu C. Chidoria bio-inspired cognitive architecture for collective robotics applica-
tions. IFAC Proceedings Volumes (IFAC-PapersOnline). 2010.
25. Ejaz N, Peterson KD, Krapp HG. An experimental platform to study the closed-loop performance of
brain-machine interfaces. Journal of Visualized Experiments: JoVE. 2011; 10(49):218228.
26. Thrun S, Bücken A. Integrating grid-based and topological maps for mobile robot navigation. In: Pro-
ceedings of the National Conference on Artificial Intelligence, 1996. pp. 944951.
27. Juliá M, Gil A, Reinoso O. A comparison of path planning strategies for autonomous exploration and
mapping of unknown environments. Autonomous Robots. 2012; 33(4):427444. doi: 10.1007/s10514-
012-9298-8
28. Chen N. An updated micromouse competition. In: Frontiers in Education Conference, 1996. pp. 1057
1059.
29. Wise RA. Brain reward circuitry: insights from unsensed incentives. Neuron. 2002; 36(2):229240. doi:
10.1016/S0896-6273(02)00965-0 PMID: 12383779
30. Olds J. Self-stimulation of the brain. Science. 1958; 127(3294):315324. doi: 10.1126/science.127.
3294.315 PMID: 13506579
31. Wallace DG, Gorny B, Whishaw IQ. Rats can track odors, other rats, and themselves: implications for
the study of spatial behavior. Behavioural Brain Research. 2002; 131(1):185192. PMID: 11844585
32. Khan AG, Sarangi M, Bhalla US. Rats track odour trails accurately using a multi-layered strategy with
near-optimal sampling. Nature Communications. 2012; 3(3):2325.
33. Beauchamp G, Yamazaki K. Chemical signalling in mice. Biochemical Society Transactions. 2003; 31
(1):147151. doi: 10.1042/bst0310147 PMID: 12546673
34. Whishaw IQ, et al. The behavior of the laboratory rat: a handbook with tests. Oxford University Press.
2004.
35. Sadik AMJ, Dhali MA, Farid HMAB, Rashid TU, Syeed A. A comprehensive and comparative study of
maze-solving techniques by implementing graph theory. In: International Conference on Artificial Intelli-
gence and Computational Intelligence (AICI), 2010. pp. 5256.
36. Abelson H, Disessa AA. Turtle geometry. MIT Press, Cambridge, Mass-London. 1981.
37. Hart PE, Nilsson NJ, Raphael B. A formal basis for the heuristic determination of minimum cost paths.
IEEE Transactions on Systems Science and Cybernetics. 1968; 4(2):100107. doi: 10.1109/TSSC.
1968.300136
38. Cully A, Clune J, Tarapore D, Mouret JB. Robots that can adapt like animals. Nature. 2015; 521
(7553):503507. doi: 10.1038/nature14422 PMID: 26017452
39. DiGiovanna J, Mahmoudi B, Fortes J, Principe JC, Sanchez JC. Coadaptive brainmachine interface
via reinforcement learning. IEEE Transactions on Biomedical Engineering. 2009; 56(1):5464. doi: 10.
1109/TBME.2008.926699 PMID: 19224719
40. Taylor DM, Tillery SIH, Schwartz AB. Direct cortical control of 3D neuroprosthetic devices. Science.
2002; 296(5574):18291832. doi: 10.1126/science.1070291 PMID: 12052948
41. Wen Y, Hou L, He L, Peterson BS, Xu D. A highly accurate symmetric optical flow based high-dimen-
sional nonlinear spatial normalization of brain images. Magnetic Resonance Imaging. 2015; 33(4):465
473. doi: 10.1016/j.mri.2015.01.013 PMID: 25620520
42. Berger TW, Hampson RE, Song D, Goonawardena A, Marmarelis VZ, Deadwyler SA. A cortical neural
prosthesis for restoring and enhancing memory. Journal of Neural Engineering. 2011; 8(4):046017. doi:
10.1088/1741-2560/8/4/046017 PMID: 21677369
43. Ramirez S, Liu X, Lin PA, Suh J, Pignatelli M, Redondo RL, et al. Creating a false memory in the hippo-
campus. Science. 2013; 341(6144):387391. doi: 10.1126/science.1239073 PMID: 23888038
44. Wu Z. The convergence of machine and biological intelligence. IEEE Intelligent Systems. 2013; 28
(5):2843. doi: 10.1109/MIS.2013.137
45. Wu Z, Pan G, Zheng N. Cyborg intelligence. IEEE Intelligent Systems. 2013; 28(5):3133.
Rat Cyborgs in Maze Solving
PLOS ONE | DOI:10.1371/journal.pone.0147754 February 9, 2016 18 / 18

Supplementary resources (7)

Data
February 2016
Data
February 2016
Data
February 2016
Data
February 2016
Data
February 2016
Data
February 2016
Data
February 2016
... For instance, a voice translation module can convert human commands into electrical stimuli in the appropriate brain areas, thereby guiding rats to exhibit expected motion behaviors (Wu et al., 2014). Target detection algorithms based on visual cues can automatically help navigation planning of rat robots based on their current position (Yu et al., 2016). ...
Article
Full-text available
Animal robots have remarkable advantages over traditional mechatronic ones in terms of energy supply, self-orientation, and natural concealment and can provide remarkable theoretical and practical values for scientific investigation, community service, military detection and other fields. Given these features, animal robots have become high-profile research objects and have recently attracted extensive attention. Herein, we have defined animal robots, reviewed the main types of animal robots, and discussed the potential developing directions. We have also detailed the mechanisms underlying the regulation of animal robots and introduced key methods for manipulating them. We have further proposed several application prospects for different types of animal robots. Finally, we have presented research directions for their further improvement.
... Modeling spike train transformation enriches our understanding of brain communication, improves the design of cognitive prostheses that restore cognitive function losses, and probably merges both brain and machine intelligence into cyborg intelligent system of the future (Wu et al., 2013Yu et al., 2016). A spike train transformation model designed for cognitive neural prostheses needs to model the nonlinearity and dynamics in spike train transformation and perform steadily in real-time application. ...
Article
Modeling spike train transformation among brain regions helps in designing a cognitive neural prosthesis that restores lost cognitive functions. Various methods analyze the nonlinear dynamic spike train transformation between two cortical areas with low computational eficiency. The application of a real-time neural prosthesis requires computational eficiency, performance stability, and better interpretation of the neural firing patterns that modulate target spike generation. We propose the binless kernel machine in the point-process framework to describe nonlinear dynamic spike train transformations. Our approach embeds the binless kernel to eficiently capture the feedforward dynamics of spike trains and maps the input spike timings into reproducing kernel Hilbert space (RKHS). An inhomogeneous Bernoulli process is designed to combine with a kernel logistic regression that operates on the binless kernel to generate an output spike train as a point process. Weights of the proposed model are estimated by maximizing the log likelihood of output spike trains in RKHS, which allows a global-optimal solution. To reduce computational complexity, we design a streaming-based clustering algorithm to extract typical and important spike train features. The cluster centers and their weights enable the visualization of the important input spike train patterns that motivate or inhibit output neuron firing. We test the proposed model on both synthetic data and real spike train data recorded from the dorsal premotor cortex and the primary motor cortex of a monkey performing a center-out task. Performances are evaluated by discrete-time rescaling Kolmogorov-Smirnov tests. Our model outperforms the existing methods with higher stability regardless of weight initialization and demonstrates higher eficiency in analyzing neural patterns from spike timing with less historical input (50%). Meanwhile, the typical spike train patterns selected according to weights are validated to encode output spike from the spike train of single-input neuron and the interaction of two input neurons.
... This article will throw light on the subject of creating a biobots control system using deep-trained neural networks. Biobotics is the newest modern direction of robotics, in which, instead of a controlled manipulator (robot), animals with implanted electrodes are used in the nervous system, particularly cockroaches [1], flying beetles [2], rats [3] and even pigeons [4]. Over the past 30 years, as a result of the development of the Moore's law, remote control chips have been significantly reduced in size. ...
Article
Full-text available
Inspired by the biobotics, we wanted to create an automatic control system using neural networks. This paper throw light on the subject of creating a biobots control system using deep-trained neural networks based on AI framework slitherin. Insect biobotic agents can be used for monitoring and assessing missions after natural disasters or earthquakes. Scientific and technological approaches created as a result of work on this project can cre-ate fundamentally new ways of interacting not only with insects, but also with other animals.
Article
Biological robot is a kind of creature controlled by human beings by applying intervention signals through control technology to regulate biological behavior. At present, the research on bio-robot mainly focuses on terrestrial mammals and insects, while the research on aquatic animal robot is less. Early studies have shown that the medial longitudinal fasciculus nucleus (NFLM) of carp midbrain was related to tail wagging, but the research has not been applied to the navigation control of the carp robot. The purpose of this study is to realize the quantitative control of the forward and steering behavior of the carp robot by NFLM electrical stimulation. Under the condition of no craniotomy, brain electrode was implanted into the NFLM of the carp midbrain, and the underwater control experiment was carried out by applying different electrical stimulation parameters. Using the ImageJ software and self-programmed, the forward motion speed and steering angle of steering motion of the carp robot before and after being stimulated were calculated. The experimental results showed for the carp robot that was induced the steering motion, the left and right steering motion of 30° to 150° could be achieved by adjusting the stimulation parameters, for the carp robot that was induced the forward motion, the speed of forward motion could be controlled to reach 100 cm/s. The research lays a foundation for the accurate control of the forward and steering motion of the aquatic animal robot.
Chapter
Are the cognitive sciences relevant for law? How do they influence legal theory and practice? Should lawyers become part-time cognitive scientists? The recent advances in the cognitive sciences have reshaped our conceptions of human decision-making and behavior. Many claim, for instance, that we can no longer view ourselves as purely rational agents equipped with free will. This change is vitally important for lawyers, who are forced to rethink the foundations of their theories and the framework of legal practice. Featuring multidisciplinary scholars from around the world, this book offers a comprehensive overview of the emerging field of law and the cognitive sciences. It develops new theories and provides often provocative insights into the relationship between the cognitive sciences and various dimensions of the law including legal philosophy and methodology, doctrinal issues, and evidence.
Article
Full-text available
Current computer vision systems, especially those using machine learning techniques are data-hungry and frequently only perform well when dealing with patterns they have seen before. As an alternative, cognitive systems have become a focus of attention for applications that involve complex visual scenes, and in which conditions may vary. In theory, cognitive applications uses current machine learning algorithms, such as deep learning, combined with cognitive abilities that can broadly generalize to many tasks. However, in practice, perceiving the environment and adapting to unforeseen changes remains elusive, especially for real time applications that has to deal with high-dimensional data processing with strictly low latency. The challenge is not only to extract meaningful information from this data, but to gain knowledge and also to discover insight to optimize the performance of the system. We envision to tackle these difficulties by bringing together the best of machine learning and human cognitive capabilities in a collaborative way. For that, we propose an approach based on a combination of Human-in-the-Loop and Knowledge Discovery in which feedback is used to discover knowledge by enabling users to interactively explore and identify useful information so the system can be continuously trained to gain previously unknown knowledge and also generate new insights to improve human decisions.
Article
Full-text available
OBJECTIVE Artificial manipulation of animal movement could offer interesting advantages and potential applications using the animal’s inherited superior sensation and mobility. Although several behavior control models have been introduced, they generally epitomize virtual reward-based training models. In this model, rats are trained multiple times so they can recall the relationship between cues and rewards. It is well known that activation of one side of the nigrostriatal pathway (NSP) in the rat induces immediate turning toward the contralateral side. However, this NSP stimulation–induced directional movement has not been used for the purpose of animal-robot navigation. In this study, the authors aimed to electrically stimulate the NSP of conscious rats to build a command-prompt rat robot. METHODS Repetitive NSP stimulation at 1-second intervals was applied via implanted electrodes to induce immediate contraversive turning movements in 7 rats in open field tests in the absence of any sensory cues or rewards. The rats were manipulated to navigate from the start arm to a target zone in either the left or right arm of a T-maze. A leftward trial was followed by a rightward trial, and each rat completed a total of 10 trials. In the control group, 7 rats were tested in the same way without NSP stimulation. The time taken to navigate the maze was compared between experimental and control groups. RESULTS All rats in the experimental group successfully reached the target area for all 70 trials in a short period of time with a short interstimulus interval (< 0.7 seconds), but only 41% of rats in the control group reached the target area and required a longer period of time to do so. The experimental group made correct directional turning movements at the intersection zone of the T-maze, taking significantly less time than the control group. No significant difference in navigation duration for the forward movements on the start and goal arms was observed between the two groups. However, the experimental group showed quick and accurate movement at the intersection zone, which made the difference in the success rate and elapsed time of tasks. CONCLUSIONS The results of this study clearly indicate that a rat-robot model based on NSP stimulation can be a practical alternative to previously reported models controlled by virtual sensory cues and rewards.
Chapter
Cognitive Vision Systems have gained significant interest from academia and industry during the past few decade, and one of the main reasons behind this is the potential of such technologies to revolutionize human life as they intend to work under complex visual scenes, adapting to a comprehensive range of unforeseen changes, and exhibiting prospective behavior. The combination of these properties aims to mimic the human capabilities and create more intelligent and efficient environments. Nevertheless, preserving the environment such as humans do still remains a challenge in cognitive systems applications due to the complexity of such process. Experts believe the starting point towards real cognitive vision systems is to establish a representation which could integrate image/video modularization and virtualization, together with information from other sources (wearable sensors, machine signals, context, etc.) and capture its knowledge. In this paper we show through a case study how Decisional DNA (DDNA), a multi-domain knowledge structure that has the Set of Experience Knowledge Structure (SOEKS) as its basis can be utilized as a comprehensive embedded knowledge representation in a Cognitive Vision System for Hazard Control (CVP-HC). The proposed application aims to ensure that workers remain safe and compliant with Health and Safety policy for use of Personal Protective Equipment (PPE) and serves as a showcase to demonstrate the representation of visual and non-visual content together as an experiential knowledge in one single structure.
Chapter
Full-text available
Data Stream represents a significant challenge for data analysis and data mining techniques because those techniques are developed based on training batch data. Classification technique that deals with data stream should have the ability for adapting its model for the new samples and forget the old ones. In this paper, we present an intensive comparison for the performance of six of popular classification techniques and focusing on the power of Adaptive Random Forest. The comparison was made based on four real medical datasets and for more reliable results, 40 other datasets were made by adding white noise to the original datasets. The experimental results showed the dominant of Adaptive Random Forest over five other techniques with high robustness against the change in data and noise.
Chapter
Cognitive Vision Systems have gained significant attention from academia and industry during the past few decades. One of the main reasons behind this interest is the potential of such technologies to revolutionize human life since they intend to work robustly under complex visual scenes (which environmental conditions may vary), adapting to a comprehensive range of unforeseen changes, and exhibiting prospective behavior. The combination of these properties supports the creation of more intelligent and efficient environments by mimicking the human capabilities. Nonetheless, preserving the environment involves gathering visual and other sensorial information and translating it into knowledge to be useful, which still remains a challenge for real time cognitive vision applications due to the complexity of such process. Experts believe the starting point is to establish a knowledge representation for Cognitive Vision technologies as a unique standard that could integrate image/video modularization and virtualization, together with information from other sources (wearable sensors, machine signals, context, etc.) and capture its knowledge. In this chapter, we present a multi-domain knowledge structure based on experience, which can be used as a comprehensive embedded knowledge representation for Cognitive Vision System, addressing the representation of visual content issue and facilitating its reuse. In addition, a successful representation and management of knowledge in cognitive systems would support communication and collaboration between humans and machines, for increased learning capabilities and enhanced decision making; this concept is a pathway towards what is called Augmented Intelligence. The implementation of such representation has been tested in a Cognitive Vision Platform for Hazard Control (CVP-HC) to address the issue of workers’ exposure to risks in industrial environments, in special for the non-use of personal protective equipment, facilitating knowledge engineering processes through a flexible and adaptable implementation.
Article
Full-text available
As robots leave the controlled environments of factories to autonomously function in more complex, natural environments, they will have to respond to the inevitable fact that they will become damaged. However, while animals can quickly adapt to a wide variety of injuries, current robots cannot “think outside the box” to find a compensatory behavior when damaged: they are limited to their pre-specified selfsensing abilities, can diagnose only anticipated failure modes, and require a pre-programmed contingency plan for every type of potential damage, an impracticality for complex robots. Here we introduce an intelligent trial and error algorithm that allows robots to adapt to damage in less than two minutes, without requiring self-diagnosis or pre-specified contingency plans. Before deployment, a robot exploits a novel algorithm to create a detailed map of the space of high-performing behaviors: This map represents the robot’s intuitions about what behaviors it can perform and their value. If the robot is damaged, it uses these intuitions to guide a trial-and-error learning algorithm that conducts intelligent experiments to rapidly discover a compensatory behavior that works in spite of the damage. Experiments reveal successful adaptations for a legged robot injured in five different ways, including damaged, broken, and missing legs, and for a robotic arm with joints broken in 14 different ways. This new technique will enable more robust, effective, autonomous robots, and suggests principles that animals may use to adapt to injury.
Article
Full-text available
A rat robot is a type of animal robots, where an animal is connected to a machine system via a brain-computer interface. Electrical stimuli can be generated by the machine system and delivered to the animal's brain to control its behavior. The sensory capacity and flexible motion ability of rat robots highlight their potential advantages over mechanical robots. However, most existing rat robots require that a human observes the environmental layout to guide navigation, which limits the applications of rat robots. This work incorporates object detection algorithms to a rat robot system to enable it to find 'human-interesting' objects, and then use these cues to guide its behaviors to perform automatic navigation. A miniature camera is mounted on the rat's back to capture the scene in front of the rat. The video is transferred via a wireless module to a computer and we develop some object detection/identification algorithms to allow objects of interest to be found. Next, we make the rat robot perform a specific motion automatically in response to a detected object, such as turning left. A single stimulus does not allow the rat to perform a motion successfully. Inspired by the fact that humans usually give a series of stimuli to a rat robot, we develop a closed-loop model that issues a stimulus sequence automatically according to the state of the rat and the objects in front of it until the rat completes the motion successfully. Thus, the rat robot, which we refer to as a rat cyborg, is able to move according to the detected objects without the need for manual operations. The object detection methods and the closed-loop stimulation model are evaluated in experiments, which demonstrate that our rat cyborg can accomplish human-specified navigation automatically.
Article
Full-text available
We investigated the ability of honeybees to learn mazes of four types: constant-turn mazes, in which the appropriate turn is always in the same direction in each decision chamber; zig-zag mazes, in which the appropriate turn is alternately left and right in successive decision chambers; irregular mazes, in which there is no readily apparent pattern to the turns; and variable irregular mazes, in which the bees were trained to learn several irregular mazes simultaneously. The bees were able to learn to navigate all four types of maze. Performance was best in the constant-turn mazes, somewhat poorer in the zig-zag mazes, poorer still in the irregular mazes, and poorest in the variable irregular mazes. These results demonstrate that bees do not navigate such mazes simply by memorizing the entire sequence of appropriate turns. Rather, performance in the various configurations depends on the existence of regularity in the structure of the maze and on the ease with which this regularity is recognized and learned.
Conference Paper
A cognitive collaborative multi-agent control architecture that addresses real-world control problems for swarms of mobile robots is proposed. The swarm's emergent behaviour is obtained by using a distributed Particle Swarm Optimization inspired algorithm. A swarm-user interface is also presented and offers a way for a human operator to interact with and guide the robotic swarm without limiting its emergent intelligence. The architecture is designed as a multi-agent system developed using JADE Framework. Three types of agents were defined: local behaviour agent, social behaviour agent and graphical user interface (GUI) agent. Each robot is associated with a pair of a local and a social behaviour agents which implement the reactive component and the interaction between the robots. The swarm forms a hierarchical structure composed of subswarms and neighbourhoods based on tasks and goals defined by the user or the swarm itself. The GUI agent is used as a link between the human expert and the swarm. The architecture was tested on a swarm of e-puck robots.
Book
This book contains a wide range of information of huge complexity on rat behavior. The book has three objectives. The first objective is to present an introduction of rat behavior. In choosing the rat as the subject species, the book has made the assumption that this species will remain, as it has in the past, the primary subject used the laboratory investigations of behavior. The second objective is to describe the organization and complexity of rat behavior. The major theme emerging from many lines of research on rat behavior is that understanding the rules of behavioral organization will be central in understanding the structural basis of behavior. The third objective is to update, as much as is possible, previous compendiums of rat behavior. Behavioral neuroscience continues to be a diverse field of research in which there remain many competing experimental methods and hypotheses. The behavioral descriptions in this book are closely tied to the laboratory methods from which they were derived, thus allowing investigators to exploit both the behavior and the methods for their own research. The first part of the book includes sections on natural history, sensory systems, motor systems, regulatory systems, development and parental behavior, social behavior, cognitive functions, and models. The second section is comprised by the major tests used by scientists interested in each domain of behavior.
Article
A cognitive collaborative multi-agent control architecture that addresses real-world control problems for swarms of mobile robots is proposed. The swarm's emergent behaviour is obtained by using a distributed Particle Swarm Optimization inspired algorithm. A swarm-user interface is also presented and offers a way for a human operator to interact with and guide the robotic swarm without limiting its emergent intelligence. The architecture is designed as a multi-agent system developed using JADE Framework. Three types of agents were defined: local behaviour agent, social behaviour agent and graphical user interface (GUI) agent. Each robot is associated with a pair of a local and a social behaviour agents which implement the reactive component and the interaction between the robots. The swarm forms a hierarchical structure composed of subswarms and neighbourhoods based on tasks and goals defined by the user or the swarm itself. The GUI agent is used as a link between the human expert and the swarm. The architecture was tested on a swarm of e-puck robots.
Article
An increasing number of functional brain imaging studies are employing computer-based virtual reality (VR) to study changes in brain activity during the performance of high-level psychological and cognitive tasks. We report the development of a VR radial arm maze that adapts for human use in a scanning environment with the same general experimental design of behavioral tasks as that has been used with remarkable effectiveness for the study of multiple memory systems in rodents. The software platform is independent of specific computer hardware and operating systems, as we aim to provide shared access to this technology by the research community. We hope that doing so will provide greater standardization of software platform and study paradigm that will reduce variability and improve the comparability of findings across studies. We report the details of the design and implementation of this platform and provide information for downloading of the system for demonstration and research applications