Systematic Benchmark for Error Recovery in Robotic Manipulation

Overview

RecoverBench examines error recovery in robotic manipulation, including slippage, pose shifts, and occlusion. The work connects structured failure analysis with multimodal recovery trajectories.

My contributions

• Co-designed a hierarchical error taxonomy and data-cleaning pipeline.
• Led interactive recovery trajectory collection with RGB-D observations and proprioception.
• Contributed data for benchmarking recovery behavior in VLA and ACT policies.

Areas & tools

• Embodied AI
• VLA / ACT
• RGB-D