HUMAN-ROBOT TEAMING STRATEGY FOR FAST TELEOPERATION OF A
LUNAR RESOURCE EXPLORATION ROVER
International Symposium on Artificial Intelligence, Robotics and Automation in Space
Virtual Conference 19–23 October 2020
L-J. Burtz1, F. Dubois2, N. Guy3
1Amanogi Corp., Japan, E-mail: firstname.lastname@example.org
2Datamapab, Japan, E-mail: email@example.com
3Astroscale Japan Inc, Japan, E-mail: firstname.lastname@example.org
We leverage the Lunar South Pole resource
exploration rover simulation environment from the
NASA Space Robotics Challenge to propose a
human-robot teaming strategy for fast teleoperation.
We explore the applicability to mission scenarios by
defining the roles and sharing of responsibilities
between the human operator and the AI,
implementing an AI running onboard the (simulated)
rover, and implementing the command and data flow
between the AI and the human operator through
algorithmic explainability and a Graphical User
1 INTRODUCTION AND MISSION
This study focuses on the first few upcoming Lunar
surface exploration missions. We are concerned with
the missions launching in 2021 to 2025 that include a
mobile rover that will prove the technology and lay
the groundwork for enabling increased capability on
each subsequent mission.
1.1 Human-Robot Teaming Rationale
In this context we prioritize robustness to unknown
environments and a deployed navigation architecture
that naturally evolves from iterative development.
We make algorithmic explainability a first-class
citizen to benefit development, commissioning on the
Lunar surface, and fast identification and resolution
of off-nominal scenarios.
This approach motivates a teaming strategy between
the Artificial Intelligence onboard the rover and the
human operator on console in Mission Control. The
teaming leverages the advantages and mitigates the
drawbacks of each actor:
AI onboard the rover:
- Advantages: access to high-framerate
high-resolution data and no latency induced by
- Drawbacks: limited computing resources (mainly
due to power and radiation constraints) and
difficult to prove high reliability, especially in
Human operator in Mission Control:
- Advantages: human cognition and an appreciation
of the entire mission context, future goals and past
achievements. Holds accountability for mission
success and is efficient at directing anomaly
- Drawbacks: prone to fatigue and some mistakes,
requires repetitive tasks to be abstracted, only
access to compressed, lower fidelity and delayed
Figure 1: elementary human-robot teaming
We define this human-robot teaming concept as a
system whose components have the following
- An AI running onboard the rover that can be fully
autonomous in the majority of scenarios.
- An AI designed as a “white box” with
explainability at its core, to enable real-time
tele-monitoring by a human operator.
- An AI that is able to gracefully fallback to human
tele-operation when encountering corner cases.
This avoids compromising on AI capability and
complexity while keeping development effort
- An operator that is trained to monitor the AI and
manually perform the tasks the AI fails at.
- An operator within a team that includes the other
rover operators, mission scientists and support
engineers in Mission Control.
- A Graphical User Interface to intuitively monitor
the AI and switch from “AI Behavior” to “Human
1.2 Enablers of Human-Robot Teaming
This teaming strategy requires assumptions that we
believe to be valid for the majority of the first Lunar
surface mobile exploration missions.
First and foremost, this concept is only possible
because of the relative proximity of the Moon to the
Earth. This concept cannot be applied to Mars rover
missions where the >6 min round trip communication
time would be prohibitive. Second, the data link
between the rover and Mission Control must be
bi-directional, continuous, and reliable: subsequently
referred to as Near Real Time (NRT) communication.
This is either available as long as the rover is near the
lander (e.g., ispace-inc Payload User Guide ) or
as long as there is line of sight for direct-to-Earth
communications (such as for the NASA VIPER rover
). These conditions will be met in the first phase
of any mission and will most likely be met for the
majority of the mission duration in the context of the
first upcoming Lunar surface exploration missions.
A key enabler of this Near Real Time concept itself is
that the bandwidth required is low (on the order of
100kbps downlink / 1kbps uplink), as we explain in
greater detail in the section 4.1 Bandwidth Reduction.
Also critical to this concept are enablers that we will
detail further in upcoming sections:
- Algorithmic explainability
- Clear separation of responsibilities between the
human operator and the AI
- Graphical User Interfaces that are responsive,
immersive and with a low psychological load to
enable continuous long duration operation
Figure 2: Overview functional diagram: Rover,
Communications and Mission Control
1.3 Leveraging the NASA Space Robotics
We implemented this teaming strategy in the
simulation environment of the NASA Space Robotics
Challenge Phase 2 .
The competition provides a Lunar South Pole virtual
environment implemented with the open-source
Gazebo9 engine and ROS framework. Simulated
four-wheeled rovers prospect for resources in the
vicinity of the lander. Rovers are all equipped with a
forward-facing stereo camera (40 cm baseline),
LiDAR (150 degree wide field of view line
scanning), tilt motor for the vision sensors, Inertial
Measurement Unit (3-axis accelerometer and gyro),
encoders on each wheel and steering arm (position
and speed), and headlights. “Scout” type rovers are
additionally equipped with a volatile sensor,
analogous to a Neutron Spectrometer.
Figure 3: 3D simulation of the lunar surface with
rover (yellow), lander (red) and probe (yellow cube)
Figure 3 shows the randomly generated Lunar terrain,
with a lander and the scout rover that it delivered to
the Lunar surface. The low-angle illumination is
representative of polar illumination conditions. Five
craters are marked as Permanently Shadowed
Regions. The environment also includes 30 locations
that are “volatile-rich” and that will trigger the
volatile sensor within a 2m range. Finally, a cube is
hidden within the environment, conceptually
representing a lost probe (similar concept to Apollo
12 astronauts visiting Surveyor 3).
Within this environment, three tasks must be
performed within 45 simulation minutes each:
- Task 1: using a scout rover, explore the
environment to report the location of the
volatile-rich regions. This task requires accurate
localization combined with an efficient exploration
- Task 2: using two rovers (an excavator rover and a
hauler rover), navigate to volatile-rich areas to
excavate regolith and drop it into the hauler’s bin.
This task requires accurate localization and precise
coordination between two rovers.
- Task 3: using a scout rover, find a lost probe
hidden in the environment and then return to the
lander (with a specific parking pose relative to the
lander) as quickly as possible. This task requires an
efficient exploration strategy and precise object
recognition and relative pose estimation.
We leverage the NASA competition for the
simulation environment as well as the relevance of
the tasks and rover concepts to upcoming missions.
Task 1 and 3 are particularly relevant to the first
prospecting missions and will therefore be the main
focus of this study. For the NASA competition itself,
as per the rules, we submitted a fully autonomous AI
with no human input. In this study, however, we
explain how we took the concept further by:
- Exposing the “thought process” of the AI (inputs,
intermediate computations, and outputs) in a
Graphical User Interface to the human operator
- Adding the capability for the human operator to
send direct commands to the rover or modify
higher level goals for the AI to follow
- Adding an interface behavior to allow the human
operator to override the AI, perform manual
actions and hand over control back to the AI
The next sections present the teaming strategy
through a design reference scenario (section 2),
specific proofs of concept (section 3) and discussion
on the path to mission implementation (section 4).
Since this human-robot teaming concept is inherently
dynamic, videos are available in a dedicated public
2 DESIGN REFERENCE SCENARIO WITH
2.1 Separation of Responsibilities and Overview of
the Design Reference Scenario
To better understand the robot-human teaming, we
present a Design Reference Scenario (based on the
Task 3 of the NASA competition). This task requires
the rover to leave the vicinity of the lander, explore
the environment to find and approach a lost probe,
precisely compute the relative distance between the
rover and the probe, then return to the lander and
precisely align with a fiducial marker on one side of
the lander (Figure 4).
Figure 5 presents the timeline of this task,
highlighting the separation of responsibilities
between the AI and the human operator as the task
progresses. In parallel, at any time that the rover is
moving, the AI behavior and human operator have
the responsibilities outlined in Figure 6.
Figure 4: Rover’s path during the four phases of the
Design Reference Scenario
2.2 Use of Behavior Trees
While Finite State Machines (FSMs) have a long
history of being used in robotics, their main
drawback is their lack of reactivity and modularity.
Behavior Trees (BTs) solve these two issues using
two-way control transfer instead of one-way control
A behavior tree can be represented as a tree structure.
Leaves are either conditions or action nodes. Other
nodes of the tree are control flow nodes. A BT can
itself take the role of an action in another BT,
contributing to modularity.
Execution of the BT occurs at a fixed time interval,
where a tick signal is generated and propagated from
parent node to child node according to the control
flow rules. A node can return three execution
, and running
finishes when the root node returns its execution
The Sequence control flow node (
symbolized by ->
executes all child nodes until one node returns failure
. If all nodes succeed, it returns success
The Fallback (also called Selector
) control flow node
symbolized by ?
) executes the child nodes until one
node returns success or running
. If all nodes fail, it
Figure 5: Timeline of general tasks that are either the responsibility of the human operator or the AI
Figure 6: Timeline of recurrent mobility tasks that are either the responsibility of the human operator or the AI
2.3 Behavior Tree Implementation of AI
We use a minimal implementation using only
Sequences and Fallback control nodes with the
possible use of state variables. This is sufficient to
generalize decision trees, finite state machines and
teleo-reactive approaches .
To improve reactivity, we try to use stateless
idempotent tasks (e.g., turnHeadlightsOn). However,
it is sometimes necessary to use nodes with memory
(state) to keep idempotence and prevent an action
from being executed repeatedly (e.g., “move 1 meter
forward” being retriggered at each tick, leading to a
Figure 7: High level behavior for the Design
Reference Scenario (NASA competition round 3)
For the mission itself, we achieve a goal-oriented
design, using the fact that BTs generalize
teleo-reactive approaches. Implicit sequences make
the design a succession of goals (post conditions)
and tasks required to achieve each goal, along with
their preconditions. This is visible in this high-level
tree for the Design Reference Scenario (Fig 7) where
three phases have been identified.
In contrast with the overall behavior, which prevents
the re-execution of an achieved goal, the subtree for
Phase 1 (Fig 8) will re-execute former tasks if a
precondition is no longer true. For example, losing
sight of the probe will retrigger the search behavior.
This is an example of reactivity creating more robust
Figure 8: Behavior for the probe search phase
2.4 Teaming with Behavior Trees
Since BTs generalize decision trees, we can create an
auto/manual handover mechanism.
Fig. 9 shows the overall BT, where the autoMode
node refers to the fully autonomous BT for the
mission in progress (for example Fig 7). We see that
the modularity of BTs is beneficial, since the details
of the autonomous behavior can be ignored at this
Figure 9: Human operator / AI behavior control
Two levels of handover are available. At a high level,
the manual/autonomous switch transfers mobility
control between the AI and the human operator. This
allows the operator to either send high-level goal
coordinates to the planner via goToGoal, or direct
low-level commands via waitForCommands (go
In both manual and automatic modes, by default a
safety check of the absence of obstacles in close
proximity to the rover (called “Auto Emergency
Stop”) is active. This low level safety check can also
be disabled by the human operator.
2.5 Navigation Teaming via Factor Graphs
The navigation capability is one of the main system
design drivers for a Lunar rover. Localization and
mapping are needed to ensure safety of the rover (not
losing communication coverage and not going into
hazardous regions) and for mission objectives (report
where interesting instrument observations have been
made and follow a systematic survey strategy). The
teaming concept allows us to select a specific
navigation strategy that leverages:
- access to high-resolution/high-frame rate sensor
input onboard the rover (but this is limited by
- access to the human cognition of the operator in
We propose a teaming strategy for Simultaneous
Localization and Mapping (SLAM) implemented via
a factor graph. This graph links together observations
(factors) at high semantic level (probe, lander,
landmarks sightings) completed by odometry at high
frequency. The resulting graph is then solved with
Multi-modal iSAM . The factor graph contains
variables representing the pose of the rover or
landmarks at different points in time, and related to
each other by factors representing knowledge about
absolute or relative positions. For example, with
odometry, we can relate to rover poses in the graph
using the pose delta in the odometry frame.
Observations of landmarks such as the lander using
the camera also create a factor between the rover
pose and the lander at observation time.
We choose a high semantic level implementation
because of the lack of features in the Lunar
environment (both in the simulation and in difficult
illumination conditions on the Moon), to reduce the
size of the graph (lower computational cost for
solving and lower bandwidth cost for downlink) and
to make the graph more human interpretable.
We provide interfaces for the human operator to
interact with the factor graph in multiple ways:
- Annotation of observations: by tagging or image
- Removal of observations: by reviewing sensor
inputs at and around the time a factor has been
added. This can be used to remove erroneous or
- Addition of observations:to add observations that
were missed or to add external observations (e.g.,
observation from an orbiter, or from a camera on
the lander) or new types of landmarks.
Figure 10: A factor graph after a few observations
(blue rectangular nodes) linking rover poses and
landmarks (red nodes)
Figure 11: Visualization of the point cloud from the
stereo camera (with visualization of the rover body
for context), to manually add or review landmarks
3 SHOWCASES: SPECIFIC PROOFS OF
3.1 Overcoming Hazardous Terrain
Obstacle detection itself uses input from a horizontal
LiDAR line scan, which is processed into higher
level primitives: segments and circles representing
discrete obstacles (rocks, slopes, the lander, etc.). The
rover's local path planning uses the Timed Elastic
Band (TEB) algorithm  to compute the optimal
path toward a given goal that avoids obstacles. For
safety, the hazard avoidance settings are deliberately
conservative, and sometimes result in the rover going
around hazardous terrain (e.g., a medium slope hill)
when it would in fact have been safe to go straight
through the hazardous terrain (e.g., over the hill).
The human operator can assess the situation more
completely (using more context: camera image,
previous experience, etc.) and decide that the most
efficient trajectory (including safety) to reach the
next goal is actually through the hazardous terrain.
The human operator therefore sends the more
efficient commands manually , before handing
over back to the AI that will resume execution within
the new context (e.g., after the hill).
Figure 12: Rover in front of a hill blocking it from
its goal (in this case the lander). Human operator
experience can determine it is safe to go over this hill
3.2 Call for Assistance
The AI has the capability to detect that it is
confronted to a new or dangerous situation. In these
cases, the behavior is set to stop the rover and switch
to human control. Once the problem is identified and
solved manually, the solution can be encoded into a
new behavior (effectively teaching the AI).
3.3 Precision and Context-aware Mobility
At the end of the Design Reference Scenario (Task 3
of the NASA challenge), the last task is to find the
fiducial marker on the lander and precisely align with
it (this task conceptually represents connecting to a
charging port or handing over samples to the lander
The AI behavior uses as input a Marker Pose node
that matches the features of a known model of the
fiducial marker with the features in the rover’s
camera image (see Appendix 1 Technology Stack) to
determine the relative pose (position and orientation)
of the rover to the fiducial marker (Figure 13).
However, in the presence of noise in the camera
(simulated ionizing radiation effects) and difficult
illumination conditions (e.g shadow of the lander) the
AI behavior sometimes struggles to perform the
In these situations, the human operator takes over
and, using the front camera as input, sends the precise
mobility commands needed for final positioning and
alignment. This concept is similar in essence to the
manual grappling of some visiting vehicles to the
International Space Station by astronauts operating
the CanadaArm (e.g SpaceX Cargo Dragon, JAXA
Figure 13: Rover aligning with fiducial marker.
(inset right) feature descriptor matching for 3D/2D
correspondence and camera pose computation
3.4 Opportunistic Detections and Science
Regarding the probe detection task, the operator can
sometimes detect the probe (or visual cues that lead
to the probe) before the AI. In these cases, the
operator can switch to “Human Override” mode and
set a goal toward the probe. Once the automatic
detection of the probe becomes stable, the operator
can give control back to the AI. The reactiveness of
Behavior Trees to new inputs means that the AI
won’t resume at the exploration task but will resume
directly at the probe approach task.
This concept is also applied for opportunistic science
where a human operator (e.g. mission scientist)
identifies an object of interest visually or through a
combination of sensor readings (e.g from the Neutron
Spectrometer). In this case, the human operator
preempts the AI behavior and selects a new goal
toward that object of interest. The object can also be
added to the factor graph. This use case is inspired by
the opportunistic science behaviors on the Mars
4 DISCUSSION: TOWARD FLIGHT
We considered the relevance of the proposed
human-robot strategy to upcoming Lunar exploration
missions from the beginning of the design phase. In
this section, we highlight some of the main
considerations in the design space: bandwidth
reduction, mission profiles and mission phases.
4.1 Bandwidth Reduction
While we did not simulate communication delays,
Quality of Service and bandwidth limitation during
this study, these considerations were included in our
design approach. Specifically, the visualizations
needed for our teaming strategy rely only on high
semantic level, low bandwidth telemetry. The
processes onboard the rover reduce the needed
bandwidth in two ways:
1. Traditional compression of images / LiDAR point
cloud, stereo camera point cloud.
2. Increasing levels of abstraction in the AI behavior
we implemented naturally lead to low data rate
Regarding the first point: The only large data needed
to be downlinked in real time to support the teaming
strategy is one NRT feed from one camera. At 2 fps,
640x480 resolution, h264 hardware compression, this
requires only about 50kbps .
The remaining large data is stored onboard the rover
and will be downlinked during non-driving phases
(pauses for battery charging, thermal balance and
operator shift changes). While these raw datasets are
not used for the real-time teaming strategy, they are
valuable for the science and engineering output of the
mission, and for better long term planning during the
Figure 14: 3D visualization of the low-bandwidth
primitives: (top) local path (green) between rocks
(red disks) and slope (yellow segments)
(bottom) lander coordinates (blue cylinder) and
probe coordinates (blue cube) from Vision AI
The second point is more interesting: the remaining
data needed for the teaming strategy consists of
high-level, discrete algorithm outputs:
- Circle center coordinate and radius for each
obstacle of the LiDAR instead of a full line scan.
- 3D coordinates of the position of the lander, rover,
marker and probe (only when detected) taken from
the stereo point cloud instead of the full point
- Behavior logic and states (several hundred bytes
The reliance on pre-processing and an active AI
onboard the rover allows for significant bandwidth
reduction during the active phases (rover moving).
This concept is a significant enabler considering the
effort and cost for the infrastructure needed to
support Lunar communications.
4.2 Mission Profiles
The teaming strategy is relevant to several mission
- Technology demonstration mission: exploration
around the lander
- Traverse to and between science stations (eg
NASA VIPER concept of operations)
- Science payload (e.g. GPR or Neutron
spectrometer) need to follow a specific grid pattern
for a systematic survey.
- Virtually all mission profiles that stay within
communication range of the lander (Payload User
Guides of CLPS lander providers show that
communication range can be expected to be within
250m/500m of the lander) or mission profiles that
include direct-to-Earth rover communications (e.g.
the NASA VIPER rover)
For any mission profiles, the teaming strategy is
1. At the start of the mission for efficient
2. During off-nominal scenarios for performing
Failure Detection Isolation and Recovery (FDIR)
First, the start of any mission carries significant risk
and lessons learned show that the shakeout of all
systems to verify that they are functioning as
designed yielded surprises. Particularly for the first
few upcoming lunar surface exploration missions, we
expect surprises related to navigation tasks due to:
- difference between the simulated environment and
lunar analogs used for development (illumination
conditions, hazard distribution),
- difference between the planned and the actual
landing site, and
- issues in the performance of some rover
subsystems (issues in any of vision, mobility,
power, thermal or communication subsystems
could require adopting a new exploration strategy).
The teaming strategy proposed, specifically the ease
of modifying and adding behaviors online during the
first hours of the mission would prove invaluable.
Second, system engineering best practices consider
off-nominal scenarios from the design phase. The
teaming strategy provides inherent explainability and
ability to monitor the inner workings of the AI
behavior. This capability creates a constant failure
detection stance and provides dedicated, well
rehearsed pathways to failure isolation and recover.
Keeping the human in the loop also makes for more
flexible reconfiguration and adaptation to degraded
performance (as long as the communication link is
The specifics of the next few upcoming lunar surface
exploration missions motivate a human-robot
teaming strategy to achieve fast teleoperation of
The teaming strategy is particularly well suited in the
context of exploration (uncertain, changing, complex
goals), high capability, high reliability, manageable
The teaming strategy is a stepping stone towards fully
autonomous missions by monitoring autonomous
behavior in-situ and collecting the treasure trove of
data to properly inform the design of a fully
autonomous concept of operations for the subsequent
mission. The teaming strategy makes the most of the
human operator’s capabilities in the first stages of a
mission while a communication link is available and
prepares the rover in the best possible way for the
more ambitious beyond communication range phases.
The teaming strategy makes for efficient iterative
development which naturally evolves into the flight
Fully manual teleoperation is too slow for short
mission durations (most concepts are less than 14
Earth days) and does not scale well to multiple rovers
and the needed mission complexity. Fully
autonomous concepts of operations risk delaying the
first mission due to the complexities of development,
testing and mission assurance.
We express our sincere gratitude to the NASA Space
Robotics Challenge team and sponsors for the
simulation environment and encouragement to
publish results that expand on the competition itself.
We thank the authors of open source software and the
ROS community that made rapid iterative
development of the concepts presented in this study
(and our submission to the NASA competition)
possible. Direct links to their code repository or
reference are in Appendix 1 Technology Stack.
 Gaines D. et al. (2016) Productivity Challenges
for Mars Rover Operations: A Case Study of Mars
Science Laboratory Operations. AAAI Int. Conf.
Automated Planning and Scheduling
 NASA Space Robotics Challenge Phase 2
lenges/space_robotics/about.html (Last Updated:
 Colledanchise M. and Ögren P. Behavior Trees in
Robotics and AI: An Introduction.
 Côté N. et al. Humans-Robots Sliding
Collaboration Control in Complex Environments
with Adjustable Autonomy.
 Cognetti, M. et al. Perception-Aware
Human-Assisted Navigation of Mobile Robots on
Persistent Trajectories. (2020) IEEE/Robotics and
 Nashed, S.B. and Biswas, J. Human-in-the-Loop
SLAM. (2018) 32nd AAAI Conference on Artificial
 Fourie D., et al. (2016) A Nonparametric Belief
Solution to the Bayes Tree
 Oriol Gasquez, (2018) Hakuto Flight Model User
Personal website, while working at
ispace-inc, with permission: https://oriol.gasquez.com
 C. Rösmann, F. Hoffmann and T. Bertram, (2017)
Integrated online trajectory planning and
optimization in distinctive topologies. Robotics and
Autonomous Systems, Vol. 88, 2017, pp. 142–153.
 Uno, K., Burtz, L.-J., Hulcelle, M. & Yoshida,
K. (2018) Qualification of a Time-of-Flight Camera
as a Hazard Detection and Avoidance Sensor for a
Moon Exploration Microrover
. Transactions of the
Japan Society for Aeronautical and Space Sciences
 Moore, T., Stouch, D. (2016). A generalized
extended kalman filter implementation for the robot
. In Intelligent autonomous systems
13 (pp. 335-348).
 Walker, J. (2018) Flight System Architecture of
the Sorato Lunar Rover. 16th International
Symposium on Artificial Intelligence, Robotics and
Automation in Space (i-SAIRAS), Madrid, Spain.
 Castano, R., et al. (2007). Oasis: Onboard
autonomous science investigation system for
opportunistic rover science
.Journal of Field
Robotics, 24(5), 379-397.
 ispace-inc team. (2020) Payload User’s Guide
Available at https://ispace-inc.com/wp-content/
1.pdf, Accessed Oct. 5th, 2020.
 Ennico-Smith, et al. (2020). The Volatiles
Investigating Polar Exploration Rover Payload.
Lunar Planetary Institute, (2326), 2898.
APPENDIX 1: TECHNOLOGY STACK
LiDAR (front-facing 2D line scan)
Merge returns into segments. Merge
short segments into circular obstacles
Based on obstacle_detector ROS
Wheel and steering arm encoders + IMU
(3-axis accelerometer and gyro)
Fusion via Extended Kalman Filter
Based on robot_localization ROS
Local path planning
Current position +
next local goal ~5 to 25m away +
Compute intermediate poses to reach
the next goal while avoiding hazards
and optimizing for shortest
Based on teb_local_planner ROS
Time synchronized image pair from
front stereo camera + camera calibration
Compute disparity and point cloud with
Based on stereo_image_proc
Monocular camera image
Compute bounding boxes for the
Lander / Marker / Probe / other Rover
Custom detector based on OpenCV
color extraction in HSV color space and
Point Cloud from Stereo processing
Bounding box from Object Detection
Matching, filtering and frame
transformations to output relative
position between object and rover
Custom heuristic for combining inputs,
rejecting noise, and providing a robust
Marker Pose for precise
alignment of rover with
Time-synchronized image pair from the
front stereo camera
bounding box of marker from Object
Precisely compute the relative camera
pose (position + orientation) to the
fiducial marker on the lander
Based on OpenCV implementation of
ORB feature extraction and description,
brute force matching, and solving PnP
2D/3D correspondence with RANSAC
and decision making
Local path planning
Vision AI, MarkerPose
human operator input
Behavior Tree for teleo-reactive
Modular re-usable and mostly stateless
see Section 2.3
Custom behavior tree library
human operator input
Limited number of factors but high
quality and high semantic level. Unified
framework for including disparate
factors (odom/visual ranging/human
Based on Multi-modal RoME.jl
implementation of iSAM  in Julia
Models and plugins by NASA SRP2
Gazebo9 engine, physics by Open
All credit goes to the team at NASA
Built from source with Python 3
RVIZ for the 3D UI
ipywidgets for the interactive dashboard UI
plot_juggler for all time series plotting (most valuable tool for development and troubleshooting)
APPENDIX 2: USER INTERFACE AND USER EXPERIENCE
(see poster below)
This module is responsible for processing visual and LiDAR inputs and detecting the objects of interest
in the environment (lander, probe...) necessary for the mission execution by the AI and human operator.
We use Behavior Trees (BT) due to their modularity, reactiveness and inherent explainability .
Behavior trees replace finite state machines using a tree description of composable behaviors made up
of sequence, fallback, condition and action nodes.
Teaming Logic Implementation for Mobility
This module is also implemented as a BT (Fig. 4 below), allowing a common emergency stop capability
(“isSafe”) in both “autoMode” (itself a behavior tree) and manual
Navigation Teaming via Factor Graphs
We use a graph-based SLAM for navigation, relying on high
semantic level observations (odometry, landmarks) .
Another layer of teaming is through collaboration on the
SLAM factor graph. The operator can annotate, add,
and even remove observations.
The user interface is crucial to effective teaming . It
allows the operator to monitor the vision outputs, the
state of the AI, and the planner output (Fig. 3), and
enables the operator to override the AI through
a manual command interface (Fig. 2).
Human-Robot Teaming Strategy for Fast Teleoperation of a Lunar Resource Exploration Rover
L-J. Burtz1, F. Dubois2 and N. Guy3,
1Amanogi Corp. (email@example.com), 2Datamaplab (firstname.lastname@example.org), 3Astroscale Japan Inc. (email@example.com).
International Symposium on Artificial Intelligence, Robotics and Automation in Space - October 19-23, 2020
Figure 1: 3D simulation of the lunar surface with rover (yellow), lander (red) and probe (yellow cube)
Introduction and Mission Context
In the context of Lunar surface exploration missions being launched within the next 5 years,
we prioritize robustness to unknown environments and a deployed navigation architecture that
naturally evolves from iterative development (Fly as you Test). We make algorithmic explainability a
first-class citizen to benefit iterative development, commissioning on the lunar surface as well as fast
identification and resolution of off-nominal scenarios. This approach, combined with lessons learned
from Mars surface exploration, motivates a teaming strategy between the Artificial Intelligence
onboard the rover and the human operators on console in Mission Control.
This study is based on the NASA Space Robotics Challenge Phase 2, a competition within a
Lunar South Pole virtual environment implemented with the open-source Gazebo 9 engine and ROS
framework . For this study, we leverage the simulation environment and the relevance of the rover
concept (mobility and sensing capabilities) and tasks (prospecting for volatiles and finding a lost probe
in the vicinity of the landing site). We present a human-robot teaming approach applicable to a mission
scenario by considering:
- The implementation of the AI onboard the (simulated) lunar rover
- The definition of the roles and sharing of responsibilities between the human operator and the AI
- The implementation of the data flow between the AI and the operator through algorithmic
explainability and a Graphical User Interface (Fig. 2 and 3)
The teaming strategy is an inherently dynamic process: videos are available at a public repository:
 Gaines D. et al. (2016) Productivity Challenges for Mars Rover Operations: A Case Study of Mars Science Laboratory
Operations. AAAI Int. Conf. Automated Planning and Scheduling
 NASA Space Robotics Challenge Phase 2.
www.nasa.gov/directorates/spacetech/centennial_challenges/space_robotics/about.html (Last Updated: Aug.12, 2019)
 Colledanchise M. and Ögren P. Behavior Trees in Robotics and AI: An Introduction. (2017) arXiv preprint,
 Côté N. et al. Humans-Robots Sliding Collaboration Control in Complex Environments with Adjustable Autonomy.
 Cognetti, M. et al. Perception-Aware Human-Assisted Navigation of Mobile Robots on Persistent Trajectories. (2020)
IEEE/Robotics and Automation Letters.
 Nashed, S.B. and Biswas, J. Human-in-the-Loop SLAM. (2018) 32nd AAAI Conference on Artificial Intelligence.
Figure 3: RVIZ-based 3D representation of rover (red mesh), odometry variance (red ellipse),
odometry path (blue arrows), stereo ranger output (lander as a blue cylinder), global path
planning goals (outward spiral as green arrows). Grey grid lines are spaced every 1m.
Graphical User Interface for Human Navigator Operator in Mission Control
Figure 2: Human operator dashboard for selecting the human-robot teaming mode (top left),
for monitoring the AI via indicators of its inputs, internal states and outputs (left panel),
and for sending manual commands (right panel)
We express our sincere gratitude to the NASA Space Robotics
Challenge team and its sponsors for the simulation environment
and encouragement to publish results that expand on the
We thank all the authors of Open Source Software and the ROS
community that made rapid iterative development possible.
Showcases: Specific Proofs of Concept
Overcoming Hazardous Terrain
The hazard avoidance settings of the local planner are deliberately conservative (to ensure safety), and
sometime result in the rover going around hazardous terrain (e.g., a medium slope hill) when it would in
fact have been safe to go straight through the hazardous terrain (e.g., over the hill). The human operator
can better appreciate the context, elect to send the more efficient commands manually , before
handing over back to the AI that will resume execution within the new context (e.g., after the hill).
Call for Assistance
The AI has the capability to detect that it is confronted to a new or dangerous situation. In these cases,
the behavior is set to stop the rover and switch to human control. Once the problem is identified and
solved manually, the solution can be encoded into a new behavior (effectively teaching the AI).
During the probe detection scenario, the operator can sometimes detect the probe (or visual cues that
lead to the probe) before the AI. In these cases, the operator can switch to “Human Override” mode and
set a goal toward the probe. Once the automatic detection of the probe becomes stable, the operator can
give control back to the AI. The reactiveness of Behavior Trees to new inputs means the AI won’t
resume at the exploration task but will resume directly at the probe approach task.
This concept is also applied for opportunistic science where an operator visually identifies an object of
interest. In this case the operator preempts the AI behavior. The object of interest can also be added to
the factor graph. This use case is inspired by the opportunistic science behaviors on the Mars rovers .
The human-robot teaming concept is defined as a
system with the following characteristics:
- An AI running onboard the rover that can be
fully autonomous in the majority of scenarios.
- An AI designed as a “white box” with
explainability at its core, to enable efficient
real-time tele-monitoring by a human operator.
- An AI that is able to gracefully fallback to
human tele-operation when encountering corner
cases. This avoids compromising on AI
capability and complexity while keeping
development effort manageable.
- An operator that is trained to monitor and
preempt the AI and manually accomplish the
tasks the AI fails at.
- An operator within a team that includes the
other rover operators, mission scientists and
support engineers in Mission Control.
- A Graphical User Interface to intuitively
monitor the AI and switch from “AI Behavior”
to “Human Override” (Fig. 2 and 3)
Figure 4: Top-level Behavior Tree
Enablers of Human Robot Teaming
- < 10 seconds round-trip communication
delays (Lunar surface exploration)
- Bi-directional, continuous, and reliable
data link with the rover
- Algorithmic explainability (“white box”)
- Clear separation of responsibilities
between the human operator and the AI
- Graphical User Interfaces that are
responsive, immersive and with a low
psychological load to enable long
duration operation (6+ hour shifts)
 Fourie D., Leonard J, Kaess M, (2016) A Nonparametric Belief Solution to the Bayes Tree, IROS
 Oriol Gasquez, (2018) Hakuto Flight Model User Interface. Personal website, while working at ispace-inc, with permission:
 C. Rösmann, F. Hoffmann and T. Bertram, (2017) Integrated online trajectory planning and optimization in distinctive
topologies. Robotics and Autonomous Systems, Vol. 88, 2017, pp. 142–153.
 Uno, K., Burtz, L.-J., Hulcelle, M. & Yoshida, K. (2018) Qualification of a Time-of-Flight Camera as a Hazard Detection
and Avoidance Sensor for a Moon Exploration Microrover. Transactions of the Japan Society for Aeronautical and Space
Sciences, Aerospace Technology 16, 619–627.
 Moore, T., & Stouch, D. (2016). A generalized extended kalman filter implementation for the robot operating system. In
Intelligent autonomous systems 13 (pp. 335-348). Springer, Cham.
 Walker, J. (2018) Flight System Architecture of the Sorato Lunar Rover. 16th International Symposium on Artificial
Intelligence, Robotics and Automation in Space (i-SAIRAS), Madrid, Spain.
 Castano, R., et al. (2007). Oasis: Onboard Autonomous Science Investigation System for Opportunistic Rover Science.
Journal of Field Robotics, 24(5), 379-397.
Human-Robot Teaming Selector
The topmost section of the UI is at the core of the Human Robot Teaming
concept. The human operator is always responsible for the rover’s actions. The
role of the human operator is to monitor the execution of the AI. If the
high-level behavior of the AI deviates from a safe course of actions or is not
the most efficient with respect to the human’s appreciation of the context and
the current situation, then selecting the “Human Override” button will pause
the AI behavior. In this case, the rover stops and all other movement related
actions come to a graceful halt. In this “Human Override” mode, the human
can take the time to:
- Review any combination of telemetry and logs
- Edit the local and global goals
- Force transitions in the Behavior Tree
- Perform mobility tele-operation
Applicable in Human Override mode: the human operator
can choose to also disable the low-level “Auto Emergency
Stop” behavior. This behavior is a standalone safeguard for
the rover: it uses the front-facing LiDAR as input to trigger
an Emergency Stop command to the wheels when any
hazard is detected less than 0.5m ahead of the rover. This
distance corresponds to the minimum safe maneuvering
distance. The main rationale for this safeguard is the time
delay between sending a movement command and receiving
the acknowledgement (several seconds). Field testing on the
Hakuto lunar rover has shown the necessity of such a
safeguard . The option to disable it is provided for
This panel visualizes the state
variables of the Behavior Tree. It
is organized by phase. Within
each phase, each button becomes
green to represent that an
intermediate goal has been
achieved. Testing has shown that
this is the best way for the human
operator to immediately grasp
which task the AI behavior is now
trying to achieve.
This panel is laid out in the same
way as the “Manual Mobility
Commands” panel to provide
intuitive understanding between
the feedback of what mobility
commands the AI is sending and
the mobility commands that the
human operator could send.
Speeds are useful to monitor both
safety (appropriately low) and
efficiency (appropriately high)
depending on the terrain and
This Navigation Panel reports the index of the
currently active goal in the global goal sequence.
Testing has shown it to be a useful indicator of
the mission progress. The escaping button
becomes orange when the rover performs a
backward maneuver because it has detected that
it is stuck (note: a manual override should be
performed before this behavior becomes
This AI Vision panel visualizes the output of the AI’s object
detector and stereo ranger. Each indicator becomes green
when both the relevant object detection is made in the image
and when sufficient stereo data points are available within
the bounding box of the object to provide a reliable relative
position estimate. This indicator is an important input to
higher level behaviors of the AI’s behavior (e.g., the lander
ranging is used to set intermediate goals when the task is to
return near the lander). In the event of faulty behavior, they
allow the human operator to differentiate between issues in
the lower-level vision AI and other causes.
This bottom panel provides additional telemetry that
is relevant to the Navigator Operator’s tasks.
Of course, in a mission context, there are many more
telemetry channels and these will be visualized on
other dedicated screens for other operators. For
example, during the Hakuto project, Gasquez et al
developed a flight-ready integrated Graphical User
Interface for the Pilot and System Manager Operators
including visualization for telemetry channels such as
Power, Communications, Thermal, and all others .
This panel provides the interface for the human operator to send either of the following:
- Low-level commands directly to the rover : “Headlight” luminosity, “Sensor Tilt” angle,
“Straight Speed”, “Turn Speed”, “Brake Force”, “Steering Arm” angle (for crab movement)
- Higher-level commands to influence the Behavior Tree itself:
(top right) Three buttons force the transition to the next phase of the mission
(top left) A view from the front camera of the rover provides situational awareness and an
intuitive interface for the human operator to set goals by clicking (yellow star) directly on the
image (short distance only, as it requires stereo data to translate the pixel coordinates into a
3D coordinates). The button below the image allows for reviewing the goal coordinates
before setting them as the next goal.
3D Visualization: Rover Model
(Red) 3D mesh of the rover: The rover itself moves in 3D space to reflect the telemetry of
the Rover Odometry. The wheels, steering arms and sensor mount also rotate to reflect
telemetry from the motor encoders. Testing has shown that this visualization helps
identify situations where the rover is stuck or slipping excessively.
An ellipse centered on the rover visualizes the axial and transversal standard deviation of
the Rover Odometry. The size of the ellipse grows with the standard deviation for each
axis. For small standard deviation values, the ellipse is orange. For standard deviations
higher than 5m the size of the ellipse is capped (for visibility) and the color becomes red.
3D Visualization: Poses as arrows
Green arrows: represent the intermediate goals. The rover will travel from one green arrow to the next.
The orientation of the arrow represents the orientation the rover should have when arriving at this goal.
Blue arrows: visualize the previous path of the rover, as computed by the Rover Odometry. An arrow is
plotted every 1s; the direction of the arrow represents the orientation of the rover at that time. In the
example image above, the blue path usually connects the green arrows, except when a hazard (eg a
rock, not shown) was detected and avoided by the local path planner.
Red arrows: the local path-planning output is visualized as a sequence of red arrows going from the
current pose of the rover to the next goal. The human operator can verify that the local path-planning is
correctly avoiding hazards and choosing the optimal way forward. The green lines represent the other
topologies considered during optimization .
3D Visualization: Object Detector
Blue cylinder: visualizes the computation of the vision AI when
detecting the lander. Each new computation will update the position of
the lander. The latest position is stored in memory to serve as a
navigation aid to the human, even when the lander is no longer in the
field of view of the cameras.
Similar simplified representations are used to visualize the vision AI
output for the detection of the probe (blue cube) and the lander’s
fiducial marker (blue square).
Human-Robot Teaming Strategy for Fast Teleoperation of a Lunar Resource Exploration Rover
L-J. Burtz1, F. Dubois2 and N. Guy3,
1Amanogi Corp. (firstname.lastname@example.org), 2Datamaplab (email@example.com), 3Astroscale Japan Inc. (firstname.lastname@example.org).
International Symposium on Artificial Intelligence, Robotics and Automation in Space - October 19-23, 2020