Register now After registration you will be able to apply for this opportunity online.
Collision-free object reaching & grasping with a legged mobile manipulator
Reaching and grasping an object of interest is a relatively simple
task that can be achieved robustly in case the object is equipped
with a simple handle and a visual marker. However, often the difficulty in the task originates from the rest of the environment.
The object may be placed in cluttered spaces with diverse obstacles
as well as dynamic entities, e.g. humans, other robots. As a result,
executing the task of reaching and grasping the object necessitates
collision-free motion control capabilities.
Keywords: Mobile Manipulation, Loco-Manipulation, Reinforcement Learning, Robot Control, Legged Robotics
While numerous model-based approaches consider collision avoidance with the environment [1], incor-
porating these capabilities on model-free RL approaches remains relatively unexplored. This project will investigate developing a RL-based motion controller for a quadrupedal manipulator that needs to
reach an object in a constrained environment. First the case of static entities in the scene (e.g. object
placed on a chair/table) will be studied and the approach will then be extended to dynamic obstacles.
The developed learnt policy will need to consider inputs from perception (e.g. pointcloud) [2]. To
that end, selecting the most suitable representation for the perception data during the training and
deployment phases will be studied. Finally the project will target a real-world demonstration of our
quadrupedal manipulator ALMA [3] for collision-free reaching and offloading an object from another
mobile robot.
While numerous model-based approaches consider collision avoidance with the environment [1], incor- porating these capabilities on model-free RL approaches remains relatively unexplored. This project will investigate developing a RL-based motion controller for a quadrupedal manipulator that needs to reach an object in a constrained environment. First the case of static entities in the scene (e.g. object placed on a chair/table) will be studied and the approach will then be extended to dynamic obstacles. The developed learnt policy will need to consider inputs from perception (e.g. pointcloud) [2]. To that end, selecting the most suitable representation for the perception data during the training and deployment phases will be studied. Finally the project will target a real-world demonstration of our quadrupedal manipulator ALMA [3] for collision-free reaching and offloading an object from another mobile robot.
- Literature research. Starting with following papers:
[1] J-R. Chiu et al. A collision-free mpc for whole-body dynamic locomotion and manipulation. In
2022 international conference on robotics and automation (ICRA), pages 4686–4693. IEEE, 2022.
[2] Takahiro Miki, Joonho Lee, Lorenz Wellhausen, and Marco Hutter. Learning to walk in confined
spaces using 3d representation, 2024.
[3] Y. Ma et al. Learning arm-assisted fall damage reduction and recovery for legged mobile ma-
nipulators. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 12149–12155, 2023.
[4] M. Mittal et al. Orbit: A unified simulation framework for interactive robot learning environments.
IEEE Robotics and Automation Letters, 8(6):3740–3747, 2023.
- Develop a pipeline in simulation
- Real-world deployment on the ALMA robot
- Literature research. Starting with following papers:
[1] J-R. Chiu et al. A collision-free mpc for whole-body dynamic locomotion and manipulation. In 2022 international conference on robotics and automation (ICRA), pages 4686–4693. IEEE, 2022.
[2] Takahiro Miki, Joonho Lee, Lorenz Wellhausen, and Marco Hutter. Learning to walk in confined spaces using 3d representation, 2024.
[3] Y. Ma et al. Learning arm-assisted fall damage reduction and recovery for legged mobile ma- nipulators. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 12149–12155, 2023.
[4] M. Mittal et al. Orbit: A unified simulation framework for interactive robot learning environments. IEEE Robotics and Automation Letters, 8(6):3740–3747, 2023.
- Develop a pipeline in simulation
- Real-world deployment on the ALMA robot
The project is suitable for motivated students with a background in robotics and control and previous experience in programming (ROS, C++/Python). During the project the student will acquire hands-
on experience on model-free RL tools [4] applied in simulation and on real robots available at RSL lab, ETH.
The project is suitable for motivated students with a background in robotics and control and previous experience in programming (ROS, C++/Python). During the project the student will acquire hands- on experience on model-free RL tools [4] applied in simulation and on real robots available at RSL lab, ETH.
Interested students should send an email to Ioannis Dadiotis (ioannisda@leggedrobotics.com) with Victor Klemm (victor.klemm@mavt.ethz.ch) in cc.
Interested students should send an email to Ioannis Dadiotis (ioannisda@leggedrobotics.com) with Victor Klemm (victor.klemm@mavt.ethz.ch) in cc.