RoboCurate: Harnessing Diversity with Action-Verified Neural Trajectory for Robot Learning
Published in arXiv preprint, 2026
This paper addresses limitations in synthetic robot training data by introducing a framework that validates action quality through simulation comparison. The method replays the predicted actions in a simulator and assesses action quality by measuring the consistency of motion between the simulator rollout and the generated video.
Key innovations include leveraging image editing to expand observational diversity and applying video transfer techniques to enhance appearance variation. The research demonstrates substantial performance gains, reporting improvements such as +70.1% on GR-1 Tabletop (300 demos) and notable results in real-world dexterous manipulation tasks compared to approaches using only real data.
Recommended citation: Seungku Kim, Suhyeok Jang, Byungjun Yoon, Dongyoung Kim, John Won, and Jinwoo Shin. (2026). "RoboCurate: Harnessing Diversity with Action-Verified Neural Trajectory for Robot Learning." arXiv preprint arXiv:2602.18742.
Download Paper