Robots Share Visions to Shorten Training
Astribot, a Shenzhen embodied AI company, has published CoMind, a framework that lets two robots trade what each one can see and the other cannot. The system is also called Be My Eye. In a test where a partition hid a pen from the robot sent to collect it, success went from zero to every attempt once the machines shared a summary of the scene.
The problem is partial observability. Two robots assembling a carton, in Astribot's own example, can have a panel blocking one robot's view of whether its partner has bonded the base, or of when to flip the box, though the partner sees all of it. CoMind sends something far smaller than a camera feed, and Pandaily reported the work and its figures.
Training happens in two stages. A context encoder compresses each robot's camera views and its task instruction into a compact representation, guided by two objectives, predicting how relevant parts of the scene will change and tracking how far a job has progressed. So the shared message carries the state of the task rather than a picture of it.
In the second stage the robots exchange those summaries in both directions. A receiving robot fuses its partner's context with its own images, instruction and joint states, then feeds the result into a pretrained vision language action model that outputs continuous movement. Each robot still generates its own actions, so control stays with the machine rather than a central server.
During joint training, the receiver's action errors travel back through the shared channel to update the sender's encoder, so each robot learns to send what its teammate needs rather than everything it notices. Astribot says that moves the encoder beyond standalone representation learning and towards the partner's action requirements.
Astribot tested the framework on real dual robot tasks, including carton assembly, fridge restocking, quilt folding, dumbbell assembly, blind box selection and the partitioned pen pick. Blind box accuracy rose from 38.9% with no sharing, and from 72.2% under another context transfer method, to 88.9%. So the gain in success arrived with less traffic, and communication volume fell by about 80%.
Astribot was founded in late 2022 and is led by founder and chief executive Lai Jie. It builds robots whose joints run on cable transmissions, such as the S1, alongside the Lumo series of robot foundation models.
The figures are the company's own and have not been independently replicated, a limit Astribot states. The team has published the technical report on its site and on arXiv, and says it intends to explore richer shared information and more flexible interaction between robots.
So two robots can now hand each other the state of a job in place of a camera feed, and the figures behind that claim come from Astribot's own test bench. What does not exist yet is a run of the framework outside the company's lab.
CoMind, also called Be My Eye, lets two robots share a compact task summary instead of raw camera feeds
The partitioned pen pick rose from zero to 100% success, and blind box accuracy from 38.9% to 88.9%
Communication volume fell by about 80% against the unshared baseline
Astribot was founded in late 2022, is led by Lai Jie, and builds the S1 robot and the Lumo foundation models
The results are Astribot's own tests and have not been independently replicated
