THE ESSENTIALS
  • CoMind shares selected task information rather than complete camera streams.
  • Reported demonstrations do not establish independently verified performance across deployments.

Sharing information for action

A November 13 QbitAI analysis describes CoMind, a multi-robot collaboration framework from Shenzhen-based 星尘智能, also identified as Astribot. The report examines how robots can exchange useful observations while making their own action decisions, rather than relying on one controller to generate every instruction.

According to the article, a context encoder selects and compresses task-relevant visual information. Each robot combines shared context with its own observations, task instructions and body state before a vision-language-action model generates continuous actions.

Demonstrations and limits

QbitAI describes six company demonstrations, including carton assembly, folding bedding and refrigerator restocking. These illustrate complementary viewpoints and coordination when a robot cannot directly see a target or its partner.

The analysis also explains action-level backpropagation: during training, an action-prediction error can reach the partner's information encoder, helping adjust what it communicates. The reported approach links message selection to the receiving robot's decisions.

These are reported research demonstrations and technical explanations, not this publication's independent tests. They do not establish reliable performance across arbitrary robot bodies, environments or commercial deployments.

Sources & context

Go to the original material. Company claims remain attributed to their sources.

01
中国机器人学会隔墙“递眼色”!成功率从0飙到100%,组队干活不烧心 – 量子位 ↗Retrieved and checked by the China AI Wire agent; evidence ledger retained internally.
Last updated November 15, 2026.Spotted an issue? Let us know ↗