Reproduce “ProDapt: Proprioceptive Adaptation using Long-term Memory Diffusion” on OpenResearch

About this paper

Federico Pizarro Bejarano, Bryson Jones, Daniel Pastor Moreno, Joseph Bowkett, Paul G. Backes, Angela P. Schoellig

ProDapt introduces a method for integrating long-term memory into diffusion-based robotic controllers using "keypoints," enabling robots to perform complex manipulation tasks solely with proprioceptive sensor data. This approach allows robots to navigate previously unmanageable obstacle courses, achieving up to 80% success in complex "Elbow" tasks where baseline methods fail, while maintaining real-time inference speeds.

Key insights

• Integrating a curated, explicit long-term memory (keypoints) into the conditioning of a diffusion model significantly enhances its ability to handle complex, proprioceptive-only navigation tasks.
• This approach allows diffusion-based controllers to overcome limitations imposed by short observation horizons without incurring the high computational costs associated with simply extending the observation window.
• By learning from past contacts, the robot can develop effective circumvention strategies for unknown obstacles, preventing repetitive collisions and enabling navigation through complex environments previously inaccessible to short-memory controllers.

Summary by alphaXiv.