Advances in Robotics: Researchers Develop New Method for Generating Training Data for Contact-Rich Manipulation Tasks

Strategic Insights -

Research published by New York University (NYU) has made significant progress in the field of robotics by developing a new method for generating training data for contact-rich manipulation tasks. The study, conducted by a team of researchers, focuses on leveraging model-based planning and optimization to create demonstrations for complex tasks that require multiple contacts. The new method, which combines a diffusion-based goal-conditioned behavior cloning approach, enables effective policy learning and zero-shot transfer to hardware for two challenging contact-rich manipulation tasks.

Key Takeaways:

  • The current limitations of teleoperation interfaces make it difficult to collect demonstrations for contact-rich manipulation tasks.
  • Model-based planning and optimization can be used to generate training data for these tasks by prioritizing demonstration consistency while maintaining solution coverage.
  • The new method developed by the researchers combines a diffusion-based goal-conditioned behavior cloning approach with sampling-based planners like rapidly exploring random tree (RRT).
  • This combination enables effective policy learning and zero-shot transfer to hardware for two challenging contact-rich manipulation tasks.
  • The study reveals that popular sampling-based planners like RRT produce demonstrations with unfavorably high entropy, motivating modifications to the data generation pipeline.
  • The research has been peer-reviewed and offers new insights into the field of robotics and automation.

Statistics:

  • The study focuses on contact-rich manipulation tasks that require complex coordination of multiple contacts.
  • The new method enables effective policy learning and zero-shot transfer to hardware for two challenging contact-rich manipulation tasks.
  • The study reveals that popular sampling-based planners like RRT produce demonstrations with unfavorably high entropy.
  • The research has been published in the IEEE Robotics and Automation Letters (Volume 10, Issue 6, 2025, pp. 6248-6255).

Sources:

  • "Should We Learn Contact-rich Manipulation Policies From Sampling-based Planners?" (2025IEEE Robotics and Automation Letters, 10(6), 6248-6255)
  • NewsRx. Reports from New York University (NYU) Describe Recent Advances in Robotics and Automation (Should We Learn Contact-rich Manipulation Policies From Sampling-based Planners?). Robotics & Machine Learning. June 30, 2025; p 358.