Reproduce “SID: Sliding into Distribution for Robust Few-Demonstration Manipulation” 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
Yicheng Ma, Wei Yu, Zhian Su, Xidan Zhang, Huixu Dong
SID treats distribution shift in few-demonstration manipulation by separating approach correction from local task execution. An object-centric motion field aligns toward demonstrated approach states, then an egocentric policy performs the task. In OOD drawer trials, SID-closed succeeds in 92% of 50 trials, versus 76% for MT3, using two demonstrations versus MT3’s 10. Recovery depends on reliable pose estimates and reachable sampled states.
Key insights
- Explicitly managing distribution shift through an online, learned motion field allows robust few-demonstration manipulation across wide spatial variations and disturbances.
- Decomposing manipulation into distinct approach and execution phases, each with specialized modules, significantly enhances robustness and data efficiency compared to single-policy approaches.
- Combining egocentric observation representations with targeted data augmentation and a confidence-based closed-loop inference pipeline enables continuous mitigation of distribution drift during complex, multi-stage tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.