Reproduce “A System for Imitation Learning of Contact-Rich Bimanual Manipulation Policies” 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
Simon Stepputtis, Maryam Bandari, Stefan Schaal, Heni Ben Amor
This research introduces a framework that enables robots to learn compliant, contact-rich bimanual manipulation policies from human demonstrations, achieving robust adaptation to spatial and temporal disturbances. The system demonstrated a 90% success rate on a complex bimanual insertion task with varying initial conditions, outperforming baselines and highlighting the crucial role of integrated spatio-temporal reasoning and force-torque sensing.
Key insights
- Dynamic spatio-temporal inference, as provided by EnBIP, is crucial for robust recovery from disturbances and adapting temporal progression in contact-rich bimanual tasks, outperforming methods with fixed phase progression.
- Force-torque sensor data is essential for achieving high success rates and robustness in contact-rich tasks, even though kinesthetic teaching is subjectively easier for human demonstrators to provide.
- The system's ability to attribute estimation errors to either spatial or temporal discrepancies through an extended state vector allows for sophisticated real-time adaptation to unexpected interactions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.