Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Learning Control Policy to Enhance Reachability of A Legged Manipulator
The project aims to improve the reachability of the quadrupedal robot by leveraging bipedal standing and arm motion.
Keywords: Legged robots, Reinforcement learning, MPC, Deep learning
With the development of control techniques, quadrupedal robots have shown robust walking behavior. Nowadays, quadrupedal robots are expected to improve their interaction ability with external environments to perform versatile tasks in society. One of the limitations is robots fetching objects from high places. Meanwhile, quadruped animals, such as dogs, actively leverage the bipedal stand skill to enlarge their reachability. Inspired by animal behavior, the quadrupedal robot could learn to stand up with two legs to increase its base height for the reachable region. This project will investigate the reinforcement learning (RL)-based controller to achieve this novel task. The method could be whole-body model-free RL or model-free RL (legs) + model-based MPC (arm).
References:
[1] Yuntao, Ma, et al. "Combining Learning-Based Locomotion Policy With Model-Based Manipulation for Legged Mobile Manipulators" RA-L 2022.
[2] Fan, Shi, et al. "Circus ANYmal: A Quadruped Learning Dexterous Manipulation with its Limbs." ICRA 2021.
With the development of control techniques, quadrupedal robots have shown robust walking behavior. Nowadays, quadrupedal robots are expected to improve their interaction ability with external environments to perform versatile tasks in society. One of the limitations is robots fetching objects from high places. Meanwhile, quadruped animals, such as dogs, actively leverage the bipedal stand skill to enlarge their reachability. Inspired by animal behavior, the quadrupedal robot could learn to stand up with two legs to increase its base height for the reachable region. This project will investigate the reinforcement learning (RL)-based controller to achieve this novel task. The method could be whole-body model-free RL or model-free RL (legs) + model-based MPC (arm).
References: [1] Yuntao, Ma, et al. "Combining Learning-Based Locomotion Policy With Model-Based Manipulation for Legged Mobile Manipulators" RA-L 2022. [2] Fan, Shi, et al. "Circus ANYmal: A Quadruped Learning Dexterous Manipulation with its Limbs." ICRA 2021.
- Literature research on RL for legged robots
- Learning framework for whole-body motion controller
- Evaluation of policy in the simulation
- Literature research on RL for legged robots - Learning framework for whole-body motion controller - Evaluation of policy in the simulation
- Creative and motivated for the topic
- Knowledge in RL, robot dynamics/kinematics
- Programming experience with Python or C++
- Creative and motivated for the topic - Knowledge in RL, robot dynamics/kinematics - Programming experience with Python or C++
- Fan Shi (fan.shi@ai.ethz.ch)
- Yuntao Ma (mayun@ethz.ch)
- Abi-Farraj Firas (fabifarraj@ethz.ch)
- Fan Shi (fan.shi@ai.ethz.ch) - Yuntao Ma (mayun@ethz.ch) - Abi-Farraj Firas (fabifarraj@ethz.ch)