Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Door Opening using Learned Foot Manipulation Skills
This project aims at enabling our quadrupedal robot, ANYmal, to perform tasks beyond locomotion and navigation. Particularly, we want to learn interaction policies that can open doors by using the robot’s legs, thereby eliminating the need for a specialized robotic arm.
Currently, most works on legged robot manipulation use an additional robotic arm payload that interacts with the door by grasping the handle [1]. However, animals such as dogs can exploit their legs for non-prehensile manipulation. Driven by this inspiration, we want to explore the usability of ANYmal’s legs for door opening with minimal hardware adaptions (for example, adding a hook).
The project will involve setting up a learning environment in a state-of-the-art simulator, NVIDIA Isaac Gym/Sim. Leveraging this simulator, the goal is to use reinforcement learning to train robust end-to-end policies for door opening with ANYmal. The problem involves multiple steps (unlocking the door and pushing the door), making it hard to specify using dense rewards. Thus, the project will also include investigating approaches to encourage exploration for robot manipulation, such as intrinsic motivation or providing meaningful priors, such as actionable maps. After training a state-based policy, we would like to approach the problem in a partially observed domain by incorporating vision.
References:
[1] Sleiman, Jean-Pierre, et al. "A unified mpc framework for whole-body dynamic locomotion and manipulation." IEEE RA-L (2021).
Currently, most works on legged robot manipulation use an additional robotic arm payload that interacts with the door by grasping the handle [1]. However, animals such as dogs can exploit their legs for non-prehensile manipulation. Driven by this inspiration, we want to explore the usability of ANYmal’s legs for door opening with minimal hardware adaptions (for example, adding a hook). The project will involve setting up a learning environment in a state-of-the-art simulator, NVIDIA Isaac Gym/Sim. Leveraging this simulator, the goal is to use reinforcement learning to train robust end-to-end policies for door opening with ANYmal. The problem involves multiple steps (unlocking the door and pushing the door), making it hard to specify using dense rewards. Thus, the project will also include investigating approaches to encourage exploration for robot manipulation, such as intrinsic motivation or providing meaningful priors, such as actionable maps. After training a state-based policy, we would like to approach the problem in a partially observed domain by incorporating vision.
References: [1] Sleiman, Jean-Pierre, et al. "A unified mpc framework for whole-body dynamic locomotion and manipulation." IEEE RA-L (2021).
- Literature research on non-prehensile manipulation and exploration in RL
- Training RL policies for door opening using NVIDIA Isaac Sim/Gym
- Incorporating vision with learned policy through teacher-student training or other approaches
- (Optional) Sim-to-real transfer of the learned policies on hardware
- Literature research on non-prehensile manipulation and exploration in RL - Training RL policies for door opening using NVIDIA Isaac Sim/Gym - Incorporating vision with learned policy through teacher-student training or other approaches - (Optional) Sim-to-real transfer of the learned policies on hardware
- Highly motivated and autonomous student interested in legged robot manipulation
- Programming experience with Python and PyTorch
- Experience with hardware is a plus
- Highly motivated and autonomous student interested in legged robot manipulation - Programming experience with Python and PyTorch - Experience with hardware is a plus