Reproduce “Physics-informed Goal-Conditioned Reinforcement Learning under Hybrid Contact Dynamics” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Vittorio Giammarino, Anastasios Manganaris, Ahmed H. Qureshi
Researchers at Purdue University extended physics-informed goal-conditioned reinforcement learning (Pi-GCRL) for contact-rich robotic manipulation by developing a contact-aware HJB residual and a hierarchical flow architecture to address issues with mode-dependent controllability. Their method enhanced performance, achieving up to 75% success in real-world pick-and-place tasks with limited offline data.
Key insights
- Full-state Eikonal regularization suffers from a fundamental 'mismatch' in hybrid systems, incorrectly applying gradient constraints to state dimensions that are locally uncontrollable but relevant for the overall task.
- The Hamilton-Jacobi-Bellman (HJB) equation's local characterization, which only constrains gradients along locally feasible directions, provides a more appropriate inductive bias for systems with hybrid contact dynamics.
- Decomposing the problem hierarchically and applying physics-informed regularization only within locally controllable state representations simplifies the learning problem and aligns the regularizer with the agent's actual influence over the system.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.