Learning compliant manipulation based on demonstration and exploration
Bachelor/Master Thesis at group TAMS
Motivation
Robots are increasingly expected to be able to deal with physical interaction tasks dexterously and flexibly. To this end, we need to develop advanced learning and control approaches to endow the robot the ability of compliant manipulation. This master thesis will focus on learning compliant manipulation based on demonstration and exploration. For example, one may consider to obtain the initial compliant manipulation policies through human demonstration. Then, exploration techniques such as reinforcement learning could be used to optimize the policies in a simulation environment. Finally, the learned manipulation policies are transferred to a real-world robot arm or hand.
Goals
- Establish a simulation environment for learning from demonstration and exploration;
- Build a learning model to encode the demonstrated multimodal data;
- Build an optimization model to refine the initial motion policies;
- Develop a controller (such as adaptive impedance controller) for task execution;
- Simulation and real-word experiments to verify the approaches.
Requirements
- Python
- ROS, Gazebo or other robot simulators
- Knowledge in Machine Learning, e.g., RL
- Knowledge in control, e.g., impedance/force control
- Knowledge in robotics, e.g., Kinematics and dynamics
Contact