Reproduce “ObstaDiff: Generalizable Diffusion Policy Learning via Obstacle-aware Representations” 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
Jiawen Wang, Kevin Yao, Khalid Jawed
ObstaDiff improves real-robot manipulation in clutter by pairing a target–obstacle–background representation with a semantic-aware encoder and diffusion-based approach policy. On 366 greenhouse executions, it achieved 75.41% overall success and an 8.20% collision rate, outperforming RGB-based ACT and Diffusion Policy baselines across target pose, obstacle layout, and appearance variations.
Key insights
- The experiments indicate that the encoder architecture matters at least as much as the added visual information: TOB inputs with a conventional ResNet encoder performed worse than the same representation with the structured encoder.
- Explicitly separating target geometry from obstacle geometry helps the policy learn safer approach directions rather than merely moving approximately toward the target.
- Modularizing approach and interaction lets the learned policy handle shared obstacle-avoidance behavior while replay libraries provide task-specific close-range motions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.