Reproduce “HACMan: Learning Hybrid Actor-Critic Maps for 6D Non-Prehensile Manipulation” on OpenResearch

About this paper

Wenxuan Zhou, Bowen Jiang, Fan Yang, Chris Paxton, David Held

HACMan introduces a reinforcement learning framework that enables robots to perform complex 6D non-prehensile manipulation tasks by learning hybrid actor-critic maps from point cloud observations. This method achieves robust generalization to unseen objects and demonstrates successful zero-shot transfer from simulation to real robots, reaching 85.4% success in diverse simulation tasks and 50% on real-world objects.

Key insights

• Spatially-grounded actions, where contact points are directly selected from object point clouds, are crucial for robust generalization and performance in diverse 6D non-prehensile manipulation.
• Temporally-abstracted actions, which encapsulate approach and retraction movements, simplify the learning process by allowing the agent to focus on crucial contact-rich phases.
• Multi-step reasoning is essential for complex 6D pose alignment tasks, as evidenced by the poor performance of a greedy policy, enabling dynamic object interactions like flipping.

Summary by alphaXiv.