COCO Matrix is building Physical AGI in the real world.
We believe that instead of enumerating the vast physical world in advance, robots should adapt and learn as they do.
COCO Matrix was founded on that belief, and in-context learning is our direction. We turn the real world into context that robots learn from.
We want robot models that see clearly, remember for longer and turn past demonstrations, corrections and experience into what the next problem needs. Data, hardware and models co-evolve around that ability.
Eventually, working with a robot becomes second nature. Anyone can be a robot's teacher, and the robot becomes your partner.
Robot models that pick up new tasks while they work, without retraining for each task.
1 of 4
Two separate runs. Shown a triangle fold (inset), a robot folds a triangle; shown a fold in half, it folds in half.
At deployment, a recording gives the pretrained model a new task to follow in its own setting. Setup needs no task-specific weight updates.
Two separate runs. Shown a triangle fold (inset), a robot folds a triangle; shown a fold in half, it folds in half.
During operation, earlier attempts and their outcomes stay in context for the current subtask. The model uses that history to adjust subsequent attempts to local constraints, such as a tight fit or a misaligned lid.
Not the left rack, not the right: on the third try the test tube goes into the blue rack.
Operators can guide the ongoing task through voice, gestures or changes to the scene. The trained model uses that guidance as it acts, allowing corrections during a motion without restarting the task.
While the robot picks up the cup, a person sets down a coaster; the cup goes there, not on the rack.
Experience recorded on another robot can provide a task prompt despite differences in body and cameras, reducing reliance on matching-body recordings. Our cross-body demonstrations so far are in simulation.
Two arms at a table sweep scraps into a dustpan; their run guides a wheeled robot, whose run guides a humanoid.
Our work on robots that learn in context, written up with the experiments behind it.
Latest: Experience is the prompt, introducing CO01. October 2026.
See all research →