We build the soft actuation, the real-time force control, and the AI that runs on them. Robin is what happens when all three are designed together: a two-armed mobile robot that works safely beside people, without a cage.Every layer of Robin is built in-house: the actuators, the real-time force control, and the physical AI. Because we own the whole stack, each layer is designed around the others. Customers get one integrated system, one team behind it, and one update path.Compliant force-controlled actuators of our own design, mechanically soft by construction rather than by software. Four US patents granted.Full-time closed-loop force control. Robin commands force directly rather than commanding a position and absorbing whatever force results. This is what the AI gets to build on.A force-aware model trained in-house, plus the fleet software that maps, navigates, records demonstrations, trains policies and runs them. One system, one update path.New skills and fixes reach the robot you already have as a software update. And because our actuators measure force at every joint, our model learns from touch as well as vision.In our own comparative testing Robin generates roughly a tenth of the impact force of a conventional rigid arm under equivalent contact conditions. That comes from the actuator design itself, which is why Robin can work without fencing.Illustrative diagram of the ratio, not test data. Both bars rise with contact speed; Robin’s stays roughly a tenth as high, and below the limit a person can safely absorb. Certified figures follow formal testing.General-purpose robot foundation models and our model solve different problems. Ours is built for one job: reliable, contact-rich industrial manipulation. That focus is how it reaches production-grade reliability with far less data and compute.Force sensing at every joint feeds the model directly, alongside vision. Insertion, hand-offs and gentle placing work because Robin reacts to what it feels.An operator teaches a new task in VR. Training and deployment run from one browser panel. New tasks go live in days.The learned policy handles contact. Behavior trees handle sequencing, retries and recovery, so the robot stays predictable and recovers from faults on its own.