Astribot’s CoMind cuts robot communication traffic by 80% in dual-task trials

Astribot announced that its CoMind (also called Be My Eye) framework enables two robots to exchange compact, task-relevant visual contexts instead of raw video streams, reducing communication volume by about 80% Pandaily AI & Tech China. The system addresses partial observability by having each robot encode its multi-camera view and instruction into a distilled representation, which the partner fuses with its own sensors before feeding a pretrained vision-language-action model for action generation.

In live tests on tasks such as restocking a fridge, assembling a carton, and selecting a marker pen across a partition, the framework delivered marked performance gains. On the partitioned pen task, success climbed from 0% without sharing to 100% with CoMind. Blind-box selection accuracy rose from 38.9% (no sharing) to 88.9% with CoMind, surpassing a competing context-transfer method that achieved 72.2%.

Astribot, founded in late 2022 and led by CEO Lai Jie, positions CoMind as a complement to higher-level task planning and plans to extend the approach to richer shared information and more flexible multi-robot interactions. Results are based on the company’s internal tests and have not been independently verified.