Morphology transfer learning with subdomain guarantees

This project aims to develop a learning model within a transfer learning framework to demonstrate kinematic path planning for redundant 7R robots. A 6R robot has an algebraic model for inverse kinematic solutions. Since the solution space is zero-dimensional (i.e., it contains a finite number of discrete solutions), the complexity of learning inverse kinematics is relatively simpler compared to redundant robots. In contrast, an nR (where n > 6) redundant robot can have a solution space of up to (n-6) dimensions, requiring strategies for redundancy resolution. A morphology transfer learning framework can help accelerate kinematic path planning for a 7R robot if a simplified 6R robot chain is identified. The advantage of transferring the learned model from the 6R robot is that it preserves the guarantees of the inverse kinematic solutions while optimizing for kinematic quality and collision avoidance by utilizing the redundant joints.

Approach

• Understand the analytical model of inverse kinematics for a non-generic (simplified) 6R robot.
• Implement a learning model for the wrist-partitioned 6R robot
• Transfer the model for a 7R robot with a 6R subchain with wrist partition (e.g., KUKA iiwa LBR, Franka Emika Panda robot)
• Compare the results (time, efficiency) with training a 7R robot from the start.

Expectation

• The student knows about implementing learning models (in PyTorch, stable_baseline3)
• The student has a mathematical background in the kinematics of serial robots.

References

 

1. Hejna, D., Pinto, L., & Abbeel, P. (2020, November). Hierarchically decoupled imitation for morphological transfer. In International Conference on Machine Learning (pp. 4159-4171). PMLR.
2. Jaquier N, Welle MC, Gams A, et al. Transfer learning in robotics: An upcoming breakthrough? A review of promises and challenges. The International Journal of Robotics Research. 2024;0(0).
3. Trabucco, B., Phielipp, M., & Berseth, G. (2022, June). Anymorph: Learning transferable policies by inferring agent morphology. In International Conference on Machine Learning (pp. 21677-21691). PMLR.
4. Manfred L. Husty, Martin Pfurner, Hans-Peter Schröcker, A new and efficient algorithm for the inverse kinematics of a general serial 6R manipulator, Mechanism, and Machine Theory, Volume 42, Issue 1, 2007, Pages 66-81