Reproduce “Instrumentation for Imitation Learning: Enhancing Training Datasets for Clothes Hanger Insertion” 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
Remko Proesmans, Thomas Lips, Francis wyffels
By integrating reflective infrared sensors into a clothes hanger, this research demonstrates that adding privileged state information through object instrumentation significantly improves the success rate of a Diffusion Policy for clothes hanger insertion. The enhanced dataset, incorporating rollouts from an instrumented expert, enables a vision-only policy to achieve comparable performance for real-world deployment.
Key insights
- Black-box imitation learning policies can autonomously recognize and prioritize instrumentation signals, sometimes overriding visual cues, without explicit programming.
- Instrumentation data enhances a policy's task awareness, leading to more robust behavior where policies tend to get 'stuck' rather than failing catastrophically (e.g., dropping the object).
- The performance benefits of instrumentation can be effectively transferred to vision-only policies by using instrumented 'expert' trajectories to enrich training datasets for 'student' policies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.