Two Robots That Cannot See Each Other Still Finish the Task: Astribot Takes Blind Collaboration From 0 to 100 Percent

Most demonstrations of robot intelligence involve one machine and one pair of eyes. Astribot, a Shenzhen embodied AI company, has published a framework that attacks the harder case: two robots that cannot see each other, cannot see the same objects, and still have to act as one pair of hands. The framework is called CoMind, or Be My Eye, and the headline result is a through-wall pen-following test that went from a 0 percent success rate to 100 percent.

The second result is arguably more useful. When one robot had to place an object into a box it could not see, relying on the partner's shared view raised accuracy from 38.9 percent, which is close to guessing, to 88.9 percent. All of this happened while the volume of information exchanged between the two robots fell by 80 percent compared with sending raw camera frames.

The mechanism is a two-stage pipeline rather than a supervisor. In the first stage each robot selects, from its own camera streams, the visual features relevant to the current task and instruction, and compresses them into a compact set of representations. In the second stage the two exchange those representations bidirectionally, and each robot feeds what it receives into its own vision-language-action model alongside its own view, the task instruction and its proprioceptive state, generating continuous action directly.

Training does the coordinating. The models learn to predict how the scene will change a moment later and to retain cues about task progress, so a partner can tell whether a step is starting or nearly finished. There is also an action-level backward path: when the blind robot misses a target, the error propagates back into the sender's encoder, teaching it which features were not distinctive enough. Over iterations the sender learns to transmit what actually changes the partner's behavior rather than what looks salient to itself.

Astribot lists six real-robot task families behind the numbers: following a pen through a barrier, placing objects into a randomly positioned box while fully occluded, assembling cardboard boxes whose faces block the view mid-flip, folding a duvet whose shape changes continuously, restocking a refrigerator where each robot sees a different shelf, and threading plates onto a dumbbell bar where the alignment tolerance is millimeters. The company says a technical report, titled Be My Eye: Learning to See Beyond a Robot's Own Perspective, was submitted to arXiv on November 9.

The work sits on top of the company's own stack: the Lumo series of base models, an embodied operating system and a tendon-driven humanoid body. Astribot positions CoMind as the next step after single-machine dexterity, with a stated plan to move from sharing information to sharing intent, so that robots anticipate each other instead of reacting. Comparisons in Chinese coverage point at other multi-robot lines such as CHORUS, LatentToM and DuoMind, with the claimed differentiator being that CoMind's shared context acts directly at the action level.

All of the figures are the company's own, from tests it designed and ran, and none has been independently reproduced. That caveat matters, because success rates in two-robot manipulation are sensitive to task setup, and a 0 percent baseline is a sign of a deliberately impossible condition rather than a normal operating point.

The direction is still the right one to watch. Factories and warehouses do not deploy one robot at a time; they deploy fleets, and the cost of coordination, in latency and in bandwidth, is one of the reasons multi-robot systems have stayed in research labs. Compressing what robots say to each other, instead of streaming everything, turns a bandwidth problem into a modeling problem. That is a far cheaper thing to scale.