Tadej Petrič’s research while affiliated with Jožef Stefan Institute and other places

What is this page?


This page lists works of an author who doesn't have a ResearchGate profile or hasn't added the works to their profile yet. It is automatically generated from public (personal) data to further our legitimate goal of comprehensive and accurate scientific recordkeeping. If you are this author and want this page removed, please let us know.

Publications (10)


Fig. 1. (a) The JSI-KneExo comprises three modular parts: control unit, left-leg exoskeleton, and right-leg exoskeleton. (b) The actuator unit incorporates air solenoid valves, PAM, pneumatic cylinder, force sensor, absolute encoder, shank and thigh lever, and a linear bearing facilitating linear PAM deflection. (c) Pink background highlights control unit components, while light blue denotes exoskeleton (leg side) components; optional insole sensors can be added.
Fig. 3. Filling compressed air into the PAM with two different pumps, the larger BD-07A-35L and the smaller BD-04A-20L. The equation fitted through measured data has the form: pm(t) = pmax(1 − e −t/k ). For the larger pump: pmax = 6.5 bar, k = 2.0713, and R 2 = 0.9921. For the smaller pump: pmax = 3.32 bar, k = 1.8302 and R 2 = 0.9969.
Fig. 4. Model-identification of the PAM DMSP-20-100N-RM-RM. In (a) PAM contraction versus pressure. In (b) PAM volume versus contraction.
Fig. 5. (a) Exoskeleton flexed at θ = 107 • , the angle where z = z 0 for ε(pm = 3.32, bar) contraction. (b) Exoskeleton at θ = 0 • for user standing.
Fig. 7. Experimental setup: EMG electrodes, reflective markers, and screen for visual feedback.

+1

Active, Quasi-Passive, Pneumatic, and Portable Knee Exoskeleton with Bidirectional Energy Flow for Efficient Air Recovery in Sit-Stand Tasks
  • Conference Paper
  • Full-text available

May 2024

·

129 Reads

·

2 Citations

·

·

·

Tadej Petrič
Download

Optimizing Robot Positioning Accuracy with Kinematic Calibration and Deflection Estimation

May 2023

·

35 Reads

To achieve higher positioning accuracy, it is common practice to calibrate the robot. An essential part of the calibration is the estimation of the kinematic parameters. Due to various nonlinear influences on the end-effector position accuracy, such as joint and link flexibility, standard methods of identifying kinematic parameters do not always give a satisfactory result. In this paper, we propose a strategy that considers deflection-dependent errors to improve the overall positioning accuracy of the robot. As joint/link deflections mainly depend on gravity, we include the compensation of gravity-induced errors in the estimation procedure. In the first step of the proposed strategy, we compute the joint position errors caused by gravity. In the next step, we apply an existing optimization method to estimate the kinematic parameters. We propose to use an optimization based on random configurations. Such an approach allows good calibration even when we want to calibrate a robot in a bounded workspace. Since calibration is generally time consuming, we investigated how the number of measured configurations influences the calibration. To evaluate the proposed method, we used a simulation of the collaborative robot Franka Emika Panda in MuJoCo.Keywordsserial manipulatorscalibrationerror compensation


Application of a Phase State System for Physical Human-Humanoid Robot Collaboration

May 2023

·

9 Reads

·

2 Citations

Collaborative robotics is one of the fastest developing but at the same time one of the most challenging fields in robotics. The reason for the latter is that a good robot collaborator should precisely understand the intentions of the human coworker. In this paper, human-robot collaboration is addressed by the application of a dynamical phase state system guided by stable heteroclinic channel networks to the motion of a humanoid robot. With this approach, a person can control the underlying dynamical system by applying forces to the grippers of the humanoid robot and this way guide the robot. The robot motions presented in this paper are lifting up the forearms and squatting, but the dynamical model can be further extended to incorporate an arbitrary amount of motions. The presented approach, therefore, provides a suitable way for the realization of efficient human-robot collaboration and should be further explored and developed.KeywordsHuman–robot interactionCollaborative roboticsPhase state systemStable heteroclinic channel networks


Examples of our work on different humanoid robotic platforms. a End-to-end vision-to-motion learning [13••]. b Learning of bimanual discrete-periodic manipulations on a humanoid robot (© [2015] IEEE. Reprinted, with permission, from [14]). c Arm synchronization for bimanual motion and obstacle avoidance and d bimanual human-robot collaboration (© [2014] IEEE. Reprinted, with permission, from [15])
Examples of adaptation of motion acquired through LbD on different platforms. a ARMAR-3 learning to wipe with visual feedback [45]. b CbI learning arm gesture motions (reprinted from [46], with permission from Elsevier). c TALOS manipulating a measuring tool. d Modification of bimanual motion on a bimanual KUKA LWR-4 platform (reprinted from [47], with permission from Elsevier). e COMAN humanoid robot full-body motion imitation (reprinted from [48], with permission from Cambridge University Press). f HOAP-3 performing full-body motion imitation — walking (© [2013] IEEE. Reprinted, with permission, from [49])
a Generalization for grasping (© [2010] IEEE. Reprinted, with permission, from [51]). b Database for generalization of reaching with both arms (reprinted from [69], with permission from Elsevier). c Generalization for periodic actions — drumming (© [2010] IEEE. Reprinted, with permission, from [53]). d Learning of CMPs and e database expansion on the KUKA LWR robot (© [2018] IEEE. Reprinted, with permission, from [73•]). Within e: (a) number of learning epochs without database — five on average. (b) Number of learning epochs with leave-one-out cross-validation — two on average. (c) Number of learning epochs through incremental database expansion — two to three on average. The numbers in the circles denote the order of learning and thus the order of database expansion (© [2018] IEEE. Reprinted, with permission, from [73•])
Three instances of human-in-the-loop intervention. a Coaching through gestures (reprinted from [46], with permission from Elsevier). b Coaching through physical interaction (© [2016] IEEE. Reprinted, with permission, from [79]). c Schematics showing quantitative sensory, and qualitative human feedback, which acts as reward for reinforcement learning (© [2018] IEEE. Reprinted, with permission, from [83])
a Humanoid robot torso in bi-manual assembly of long poles interaction (© [2015] IEEE. Reprinted, with permission, from [38]). b Humanoid robot during human robot cooperation to place the table cloth. The norm of the relative error in subsequent learning cycles is shown in the graph (reprinted by permission from Springer Nature from [87]). c The initial demonstration of wiping policy on a cylindrically shaped glass (top) and the humanoid robot while practicing the glass wiping policy on the oval-shaped glass (bottom) and d comparison of cost function evolution for AILC and hybrid AILC-RL scheme in a bar chart (© [2017] IEEE. Reprinted, with permission, from [88])
Manipulation Learning on Humanoid Robots

September 2022

·

249 Reads

·

7 Citations

Current Robotics Reports

Purpose of Review The ability to autonomously manipulate the physical world is the key capability needed to fulfill the potential of cognitive robots. Humanoid robots, which offer very rich sensorimotor capabilities, have made giant leaps in their manipulation capabilities in recent years. Due to their similarity to humans, the progress can be partially attributed to the learning by demonstration paradigm. Supplemented by the autonomous learning methods to refine the demonstrated manipulation actions, humanoid robots can effectively learn new manipulation skills. In this paper we present continuous effort by our research group to advance the manipulation capabilities of humanoid robots and bring them to autonomously act in an unstructured world. Recent Findings The paper details progress in the area of humanoid robot learning, ranging from trajectory imitation, motion adaptation in order to maintain feasibility and stability, and learning of dynamics to statistical generalization of actions, autonomous learning, and end-to-end vision-to-action learning that exploits deep neural networks. Summary With the focus on manipulation, the presented research provides the means to overcome the complexity behind the problem of engineering manipulation skills on robots, especially humanoid robots where programming by demonstration is most effective.


Modular quasi-passive mechanism for energy storage applications: towards lightweight high-performance exoskeleton

December 2021

·

123 Reads

·

6 Citations

Quasi-passive exoskeletons have emerged as a solution that avoids the high energy requirements that negatively affect the efficiency of exoskeletons. These exoskeletons do not deliver positive mechanical work to the joint, but accumulate and deliver energy in a viscoelastic element that is actively placed or removed parallel to the user's muscles. %One of the crucial aspects of the exoskeleton is the power-to-weight ratio of the actuator, as this inescapably affects the weight and thus the portability of the entire system. Previous research has investigated different strategies, mostly based on energy harvesting with compliant elements for mechanical energy storage and locking mechanisms to turn the elements on and off. This paper seeks to address the problem of bulky actuators in quasi-passive exoskeletons by experimentally evaluating the proposed quasi-passive mechanism consisting of a pneumatic cylinder, acting as an elastic element, and a solenoid valve replacing a mechanical clutch. The main advantage of the proposed mechanism is that the elastic element can be turned on and off at any time and position, with a high switching frequency, which improves the possibility to harvest the energy. Our aim was to investigate whether, with the proposed actuation, it is possible to achieve the force that has previously been shown to have a positive effect on ankle plantar flexion. Furthermore, we analyzed how the timely activation of the solenoid valve affects the force characteristics. We built the testbed equipped with sensors that allow measurements of torque, pressure, and angular deflection. The obtained measurements are in line with the theoretical model, where the force achieved is within the required range. In addition, it was shown that the time for the actuator to switch off from the peak force is about 100 ms, without major air leaks and energy bursts. The presented results highlight that the actuation of this type is a good candidate for designing a lightweight high-performance quasi-passive exoskeleton.


Exoskeleton Control Based on Network of Stable Heteroclinic Channels (SHC) Combined with Gaussian Mixture Models (GMM)

January 2021

·

22 Reads

One of the major causes of disability and sick day leaves is lower back pain. Hence it can result in a decreased life quality and lower industrial productivity. One of the possible solutions to lower back pain could be the use of exoskeletons, which would reduce the spinal loading. One of such solutions is a quasi-passive spinal exoskeleton that engages and disengages the passive support depending on the movements performed by the user. This enables the spinal support for the user when lifting a heavy load and for all other tasks, the user motion is unobstructed. To achieve autonomous clutch activation, the main challenge is to properly classify the beginning of each motion. In this paper, we proposed a novel control method that uses Gaussian Mixture Models (GMM) for movement classifiers and a network of Stable Heteroclinic Channels (SHC) for designing a phase-state-machine. Integrating GMM into the SHC network enables a fast and reliable control of the clutch mechanism of the quasi-passive spinal exoskeleton. The control system capabilities were demonstrated in an experiment with a male subject wearing the quasi-passive exoskeleton while executing three different movements representative for an industrial working environment: walking, standing, and lifting.


Maximizing the End-Effector Cartesian Stiffness Range for Kinematic Redundant Robot with Compliance

June 2020

·

46 Reads

·

4 Citations

·

·

·

[...]

·

Tadej Petrič

Compliant robots with constant joint stiffness (Serial Elastic Actuators - SEA), on the contrary to ones with variable joint stiffness (Variable Stiffness Actuators – VSA), have limited capabilities for modulating robot mechanical impedance in the interaction task. However, in the case of kinematic redundancy in specific tasks, robots can exploit the null space to adjust End-Effector (EE) Cartesian stiffness. Thus, prior knowledge of the task path or the operational workspace can be used to pre-compute joint stiffness that can enable maximal ratio between maximal and minimal stiffness of the robot’s EE during the task execution, and therefore shape achievable EE stiffness to best fit the task execution. In that light, this paper elaborates on the preselection of joint stiffnesses which influences the achievable robot’s Cartesian stiffness in a specific task. Besides optimizing the available operational EE stiffness, by pre-computed joint stiffness values, the robot will be able to adapt better to specific tasks and provide a better framework for safe and efficient physical human-robot interaction. The paper presents an approach to the selection of predefined joint stiffness values of the 7-DOFs KUKA LWR, where joint stiffness is achieved/emulated with torque feedback. In the simulation experiments, the approach is depicted in the preselection of two joint stiffness values within the prescribed range, while other joint stiffness is set constant.


Combining Virtual and Physical Guides for Autonomous In-Contact Path Adaptation

June 2020

·

15 Reads

Several approaches exist for learning and control of robot behaviors in physical human-robot interaction (PHRI) scenarios. One of these is the approach based on virtual guides which actively helps to guide the user. Such a system enables guiding users towards preferred movement directions or prevents them to enter into a prohibited zone. Despite being shown that such a framework works well in physical contact with humans, the efficient interaction with the environment is still limited. Within the virtual guide framework, the environment is considered as a physical guide, for example, a table is a plane that prevents the robot to penetrate through. To mitigate these limits we introduce and evaluate the means of autonomous path adaptation through interaction with physical guides, which essentially means merging virtual and physical guides. The virtual guide framework was extended by introducing an algorithm which partially modifies the virtual guides online. The path updates are now based on the interactive force measurements and essentially improves the virtual guides to match them with the actual physical guides.


Generation of Smooth Cartesian Paths Using Radial Basis Functions

June 2020

·

27 Reads

·

3 Citations

In this paper, we consider the problem of generating smooth Cartesian paths for robots passing through a sequence of waypoints. For interpolation between waypoints we propose to use radial basis functions (RBF). First, we describe RBF based on Gaussian kernel functions and how the weights are calculated. The path generation considers also boundary conditions for velocity and accelerations. Then we present how RBF parameters influence the shape of the generated path. The proposed RBF method is compared with paths generated by a spline and linear interpolation. The results demonstrate the advantages of the proposed method, which is offering a good alternative to generate smooth Cartesian paths.


Gaussian Mixture Models for Control of Quasi-Passive Spinal Exoskeletons

May 2020

·

279 Reads

·

29 Citations

Research and development of active and passive exoskeletons for preventing work related injuries has steadily increased in the last decade. Recently, new types of quasi-passive designs have been emerging. These exoskeletons use passive viscoelastic elements, such as springs and dampers, to provide support to the user, while using small actuators only to change the level of support or to disengage the passive elements. Control of such devices is still largely unexplored, especially the algorithms that predict the movement of the user, to take maximum advantage of the passive viscoelastic elements. To address this issue, we developed a new control scheme consisting of Gaussian mixture models (GMM) in combination with a state machine controller to identify and classify the movement of the user as early as possible and thus provide a timely control output for the quasi-passive spinal exoskeleton. In a leave-one-out cross-validation procedure, the overall accuracy for providing support to the user was 86.72±0.86% (mean ± s.d.) with a sensitivity and specificity of 97.46±2.09% and 83.15±0.85% respectively. The results of this study indicate that our approach is a promising tool for the control of quasi-passive spinal exoskeletons.

Citations (6)


... Specifically, such bio-inspired controllers make sense in high-dimensional systems where predictive forward models are unavailable. Examples include compliant robots [25,26] in complex, dynamic environments [27][28][29]. The current alternatives for systems like these are model-based controllers and high-dimensional "black-box" controllers. ...

Reference:

Stable Heteroclinic Channel-Based Movement Primitives: Tuning Trajectories Using Saddle Parameters
Application of a Phase State System for Physical Human-Humanoid Robot Collaboration
  • Citing Chapter
  • May 2023

... There are usually three approaches to changing stiffness in electromechanical VSAs. These include 1) Changing the spring preload [3,4], 2) Changing the transmission ratio between the compliant element and the load [5], and 3) Changing the physical properties of the elastic element [6][7][8]. One solution to reduce the weight and size of actuators in exoskeletons and prostheses is to use quasi-passive mechanisms, as it has been shown that negative limb work can be accumulated and reused in the elastic element to reduce the metabolic cost of walking [9], running [10], or jumping. ...

Modular quasi-passive mechanism for energy storage applications: towards lightweight high-performance exoskeleton

... The experiment emulated a typical assembly task where the interaction force was generated by the interaction between an object in the gripper and the environment. The parameters of the robot platform can be found in [30]. All values for joint stiffness were set to 400 Nm/rad. ...

Maximizing the End-Effector Cartesian Stiffness Range for Kinematic Redundant Robot with Compliance
  • Citing Chapter
  • June 2020

... Quasi-passive lumbar support exoskeletons primarily utilize passive viscoelastic elements (such as springs and dampers) to provide support, while small actuators change the support level or disengage from passive elements [40]. A hybrid exoskeleton (HExo) designed by Le et al. combines a passive upper limb exoskeleton with an active lowerback exoskeleton, driven by battery-powered servo motors [41] (Figure 11a). ...

Gaussian Mixture Models for Control of Quasi-Passive Spinal Exoskeletons