China’s ‘Doctor Octopus’ robots get a brain that can learn across bodies

- Feagine Robotics just dropped Fi0, a cross-embodiment foundation model that keeps its brain when you swap the body — and they're stress-testing it on three tendon-driven soft manipulators (A01 through A03) ranging from one flexible segment to a 50-centimeter, 6+1-DOF arm. The core idea: instead of retraining AI every time you bolt on different hardware, Fi0 ingests the robot's morphology, sensing, and actuation state through what they call an Embodiment Graph, then figures out how to execute the task with whatever limbs it's got. One human demonstration gets parsed for objects, sequence, and desired end state — no parameter updates needed. Soft robots make a brutal testbed because they bend continuously and reshape mid-task, so the AI can't lean on fixed joint angles. The bigger play challenges the humanoid-everything assumption: maybe the future isn't one universal chassis but a family of specialized machines — rigid arms for factories, soft manipulators for delicate work — all sharing one portable intelligence layer. Fi0 is early-stage and the claims need real-world validation, but if cross-embodiment learning scales, accumulated robot intelligence stops dying every time the hardware changes. Mars University would call this adequate. Pluto Uplink taught us to call it research.