HARP-Bench tests robot manipulation policies with people working in the same space

Researchers led by Qinhong Zhou have released HARP-Bench, a benchmark for robot manipulation policies in workspaces shared with active people, in a paper accepted at CoRL 2026.

Most benchmarks assume the robot is the only one acting, the authors write, but people nearby can hint at a goal, do part of the task, move objects out of reach for a while or limit where the robot can safely move. HARP-Bench sorts this into four conditions, Cueing, Contribution, Interruption and Co-presence, built from everyday tasks with versions with and without a person, and includes a pipeline to generate demonstrations.

Testing vision-language-action models, imitation learning methods and heuristic policies, the team found that the presence of people challenged current policies, with manipulation and recovery the main remaining problems. Training on HARP-Bench demonstrations improved results in simulation, and in a real-world test on one family of tasks it helped when a person contributed, but not when a person interrupted.