Reproduce “Efficient learning of goal-oriented push-grasping synergy in clutter” 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
Kechun Xu, Hongxiang Yu, Qianen Lai, Yue Wang, Rong Xiong
Researchers at Zhejiang University developed a hierarchical reinforcement learning framework for goal-oriented push-grasping in cluttered environments. Their method, which employs goal-conditioned relabeling and data-driven dense rewards via adversarial training, achieved high sample efficiency and robust sim-to-real transfer, attaining a 90.2% goal grasp success rate in challenging simulated scenarios and 86.6% in real-world tests.
Key insights
- Goal-conditioned relabeling of experiences significantly improves sample efficiency by turning what would be 'failed' attempts (grasping a non-goal object) into valuable learning opportunities for a relabeled goal.
- An inverse adversarial training approach, where a pre-trained grasping policy acts as a discriminator to guide the pushing policy, effectively provides dense, data-driven rewards for push actions, directly linking them to the ultimate goal of improving future graspability.
- Alternating training of both push and grasp policies is crucial for refining their coordination, mitigating potential distribution mismatch, and achieving robust performance in highly cluttered environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.