Researchers from UC Davis, Microsoft Research and Analog Devices tested DuoMind, a system that lets two independently controlled robot arms exchange short updates while completing simulated manipulation tasks. It matched or beat a standalone policy on most tasks across two benchmarks, but some results were close or worse and the work has not been tested on physical robots.
When two arms share a workspace, camera views do not always tell each robot what the other intends to do. Both can choose the same target or begin a joint action at different times. DuoMind gives each arm its own planner and action model rather than relying on one controller to direct the pair.
Each planner uses a vision-language model (VLM) to read the task and camera images. It sends its own action model a short instruction, then tells the other robot its intended move, expected next subgoal and relevant task information, such as an object’s location. A separate vision-language-action model (VLA) turns the instruction into movement. This division lets the planner handle coordination while the action model controls its own arm.
The team introduced RoboPoly, a simulator benchmark with seven tasks for two independently controlled Franka robot arms, including passing a bag so a partner can hang it and jointly lifting a pot. They also adapted eight RoboTwin tasks to separate policies for each arm. The authors evaluated each task over 400 simulation rollouts.
In RoboPoly’s Hang Bag task, DuoMind with a fine-tuned π0.5 action model reached a 78% success rate, compared with 52.25% for π0.5 alone. In Cook Pot, the rates were 39.25% and 1%, respectively. Gains were not universal: on RoboTwin’s Put Object Cabinet task, π0.5 alone scored 46.5%, while DuoMind scored 43.25%.
When the researchers disabled messages between agents, their tests showed more conflicts and out-of-sync actions, supporting a role for communication in this setup. The study does not show whether DuoMind works with larger robot groups.
Related coverage: A separate simulated approach inferred a robot partner’s movement limits from its behavior.
Sources
- Hanchu Zhou et al., “DuoMind: Enabling Distributed Multi-Robot Coordination with Semantic Communication”, arXiv:2610.02161v1, posted October 1, 2026.
- DuoMind code and RoboPoly benchmark repository.
- Illustrative image: Shixart1985 (photo credit: Nenad Stojković), Wikimedia Commons, licensed under CC BY 2.0. Single-arm laboratory image, not from the DuoMind study; no modifications.
