Reproduce “Data Augmentation for Manipulation” on OpenResearch

About this paper

Peter Mitrano, Dmitry Berenson

A framework for data augmentation in robotic manipulation is introduced, designed to generate physically plausible and diverse examples for deep learning models, leveraging principles of validity, relevance, and diversity. This approach led to a 14% reduction in prediction error for dynamics learning and an increase from 27% to 50% in real-robot task success rates for constraint checking.

Key insights

• Effective data augmentation for robotic manipulation requires explicit consideration of physical validity and task relevance, unlike domain-agnostic methods.
• An optimization framework leveraging rigid-body transformations can generate diverse and physically plausible augmentations for trajectories of object poses and velocities.
• The occupancy objective is paramount for preventing physically impossible augmented states, while the delta minimum distance objective helps retain task-relevant near-contact interactions.

Summary by alphaXiv.