Astribot unveils CoMind framework for robots that cooperate out of sight

Astribot, a Shenzhen-based embodied-AI company known for tendon-driven robots and its own embodied models, has released a multi-robot coordination framework called CoMind, subtitled Be My Eye. The work shifts its focus from what a single robot can do to how several robots work together.

The problem it targets is concrete. Central schedulers struggle to keep up with fine-grained manipulation, while streaming full camera views between robots floods the link with redundant pixels and burdens both communication and compute.

CoMind's answer is decentralized, proactive sharing. Each robot can push information to its partner and still generate its own actions from the context it receives, removing the dependence on one controller that issues every command, and it sends extracted visual features and manipulation states rather than raw video.

The stress tests are unusually blunt. Two robots that cannot see each other across a barrier completed a pen handover with success rising from 0% to 100%. In a blind-box placement task, accuracy went from 38.9% with no sharing to 72.2% with another context-transmission scheme, and to 88.9% with CoMind, which also cut communication volume by about 80%.

Astribot demonstrated six real-robot tasks: the barrier handover, blind-box placement, cardboard box assembly, duvet folding and storage, fridge restocking and dumbbell assembly, covering occlusion, deformable objects and precision fitting.

The method runs in two stages. A context encoder filters and compresses multi-camera observations while learning to predict future visual change and track task progress; then both robots exchange context and feed it, alongside their own vision, instructions and proprioception, into a vision-language-action model that emits continuous actions.

A second design, action-level backpropagation, ties communication to outcomes. If the receiving robot reaches for the wrong spot, the action-prediction error flows back into the sender's information encoder during training, pushing it to extract the features that actually help, so talking is optimized for the partner's next move rather than for describing the scene.

The direction is becoming a research front. Qbitai notes that a Stanford team proposed CHORUS in 2026, letting different robots share one vision-language-action policy and cooperate from independent observations without explicit runtime communication. Both lines point at the same limit. A single robot is bounded by its viewpoint, mechanics and reach, and coordination is what extends that boundary.