In simulations described in an arXiv preprint posted October 1, a helper robot inferred a partner's movement limits by watching that robot carry an object with a demonstrator, then took the demonstrator's place on a new obstacle route. The task tests a hidden coordination problem: one robot may compensate for a partner's restriction, obscuring what that partner can do.
The benchmark uses three settings: two agents carrying a rod in a 2D model, fixed-base dual-UR5 arms carrying a beam, and mobile dual-UR5 arms carrying a beam. In each, the helper first observes obstacle-free joint-carry demonstrations from a constrained robot and a demonstrator. For the test, the helper replaces the demonstrator and must carry with the same constrained robot on a new route blocked by obstacles.
For each possible physical limit, the method checks whether it explains both robots' actions. That distinction matters because an arm's missing movement can mean either that its joint cannot move that way or that the task did not require it. The demonstrator's adjustments add evidence about the constrained robot's capabilities. A planner then uses the inferred limit to avoid coordinated moves the partner could not execute.
On the new obstacle routes, the method completed 92.6% of 2D-rod tasks, 85.2% of fixed-base dual-arm tasks and 63.0% of mobile dual-arm tasks. A reference planner supplied with the true physical limits reached 96.3%, 88.9% and 70.4%, respectively. The authors report mean success rates and standard errors over three random seeds; the mobile-arm setting was the hardest and remained furthest from that reference.
The result depends on information that deployed teams may not have: the helper is assumed to know what the robots are trying to do, how they usually choose moves and how those moves affect the shared object. It also receives full observations and searches a finite list of possible limits. Because the study tests simulations only, it does not establish that a robot can diagnose hardware problems in the field. Physical tests with incomplete observations are still needed; the authors say benchmark code and evaluation scripts will be released.
Sources
- Suyu Ye et al., “Watch, Infer, Coordinate: Inferring Robot Partner Constraints for Zero-Shot Coordination”, arXiv:2610.02170v1, posted October 1, 2026.
- Illustrative photograph, not from the study: Dzikra Muhammad Imtiyaz, “Robot Lengan Industri UR5e.jpg”, Wikimedia Commons, licensed under CC BY-SA 4.0. Wikimedia serves the 1280-pixel thumbnail; no crop or other edits.
