Reproduce “Working Backwards: Learning to Place by Picking” 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
Oliver Limoyo, Abhisek Konar, Trevor Ablett, Jonathan Kelly, Francois R. Hogan, Gregory Dudek
Limoyo et al. from Samsung AI Centre and the University of Toronto developed Placing via Picking (PvP), a self-supervised method that autonomously collects expert demonstrations for complex robotic object placement tasks by reversing successful picking trajectories. The approach enables training vision-based policies that achieve 87.78% success in dish rack loading, outperforming policies trained via kinesthetic teaching by approximately 20%.
Key insights
- Reversing successful picking trajectories can generate high-quality expert demonstrations for object placement without human intervention.
- The combination of compliant control and tactile sensing is essential for achieving robustness in autonomous data collection for contact-constrained manipulation tasks.
- Noise augmentation during data collection, particularly when combined with multimodal action representations, significantly enhances policy performance and generalization capabilities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.