Reproduce “TRANSIC: Sim-to-Real Policy Transfer by Learning from Online Correction” on OpenResearch

About this paper

Yunfan Jiang, Chen Wang, Ruohan Zhang, Jiajun Wu, Li Fei-Fei

TRANSIC uses human corrections to adapt a simulation-trained policy for real-world manipulation. During data collection, an operator intervenes when the robot behaves abnormally; a residual policy learns corrections, and a gate selects when to apply them. Across four assembly tasks, it averages 81.25% success, versus 46.25% for IWR. The method depends on a base policy and is demonstrated in single-arm tasks.

Key insights

• Human online correction provides a domain-agnostic way to implicitly address multiple coexisting sim-to-real gaps (perception, dynamics, embodiment, controller) in a unified manner.
• Learning a separate residual policy from human feedback, rather than fine-tuning the base policy, prevents catastrophic forgetting and allows for data-efficient adaptation to real-world discrepancies.
• Using 3D point clouds as observations and distilling high-level end-effector actions to low-level joint positions during simulation training are crucial for robust policy transfer in manipulation.

Summary by alphaXiv.