OS3

The design premise: the VLA has full control of execution. It consumes camera observations and proprioceptive state, streams action chunks in real time, and commands the full platform, arms, head, base, and lift, as a single action space with no hand-offs between subsystems. The execution layer does not reinterpret model output. Its responsibility is physical: realizing commanded motion smoothly and safely on the hardware, at every control step.

System overview

A planner model reasons over the scene and selects what happens next; it is never inside the servo loop. The VLA executes, commanding every joint of the machine continuously. For contact-rich phases, folding, grasping, handovers, a dedicated manipulation policy assumes control of the arms while the same physics layer continues to execute underneath. The planner proposes, the body delivers, and the interface between them is where motion quality is determined.

      VISION-LANGUAGE-ACTION MODEL
   streams time-indexed action chunks
          for the whole body
                   │
                   ▼
   ┌─────────────────────────────────┐
   │ CONTINUOUS STREAMING            │
   │ handoffs splice from executed   │
   │ state · one continuous motion   │
   └────────────────┬────────────────┘
                    ▼
   ┌─────────────────────────────────┐
   │ PHYSICS LAYER                   │
   │ gravity + friction feedforward  │
   │ trajectory shaping · soft gains │
   └────────────────┬────────────────┘
                    ▼
   ┌─────────────────────────────────┐
   │ TORQUE-LEVEL JOINT COMMANDS     │
   │ stiffness · damping · position  │
   │ velocity · torque, per joint    │
   └────────────────┬────────────────┘
                    ▼
      two arms · head · base · lift

The plant

The platform is bimanual: 6+1 quasi-direct-drive joints per arm, backdrivable and torque-dense, on a dedicated CAN FD bus per side. A 2-DoF pan-tilt head provides up-down and left-right gaze. A holonomic omnidirectional base, commanded in body-frame twist, translates and rotates in any direction while carrying the full body, and a motorized CANopen vertical lift travels with it, extending the workspace from the floor to top-shelf height during base motion. Every joint accepts a full-state command on every cycle: stiffness, damping, position, velocity, and feedforward torque. The result is torque-level authority over the entire machine rather than position setpoints into a closed controller.

System identification

The controller carries a model of the machine it drives: full rigid-body dynamics of our own geometry. Gravity enters the dynamics as a configuration-dependent torque at every joint, so the model evaluates the gravity load of each arm in its current pose, with the mass distribution identified on the physical robot through static-hold calibration. Friction is treated with the same rigor: the static component that must break before motion and the velocity-dependent component that grows with speed differ per joint and per direction, so both are swept on the physical arms and fitted empirically. No parameter is taken from a datasheet. An accurate feedforward model means the servos never work against the robot's own weight, which simplifies every stage downstream.

Impedance control, feedforward first

Impedance control renders each joint as a virtual spring-damper: stiffness acts as the spring, damping as the dissipative element. A common approach raises stiffness until tracking becomes acceptable, at the cost of rigidity and hazardous contact behavior. We take the opposite approach. Gravity, friction, and velocity feedforward carry the load as torque, allowing the springs to stay soft: a stiffness gradient that relaxes from shoulder to wrist, with additional software damping beyond what the wire protocol exposes. The resulting compliance is directly observable: the arm holds posture with minimal effort, contact produces yielding rather than impact, and every joint remains backdrivable by hand.

Trajectory smoothness

Streaming chunked actions in real time is standard practice across current VLA systems. Motion quality is decided by what happens between the chunk and the motor. Smoothness is a property of the derivative chain, position, velocity, acceleration, jerk: a trajectory can be correct in position and still excite the structure if its higher derivatives are unbounded. The layer therefore shapes all of them. Chunk handoffs splice from the last executed state, the motion the arm actually performed rather than the last command issued, and trajectory shaping bounds velocity and acceleration before actuation, so discontinuous model output cannot produce discontinuous motion. Execution remains a single continuous trajectory while inference proceeds; the jerk measurements below quantify the effect.

Measured results

Every control change is validated on the physical system against recorded trajectories. Identical robot, identical tasks, before and after the layer described above:

A unified control path

The teleoperation setup used for data collection runs the identical stack the model runs: the same feedforward dynamics, the same trajectory shaping, the same cadence. Demonstrations and deployment therefore share one physics, and the data the model learns from behaves exactly like the body it will drive. Every run additionally produces a timestamped execution trace, commanded, sent, and measured, stamped with the exact control configuration that generated it, enabling exact replay for diagnosis rather than reconstruction from logs.

Safety

Motion begins only when a person initiates it, from the rear screen or the app. A stopped motion process remains stopped until inspected; no motion process restarts autonomously. Watchdogs halt the base when commands go stale, and every stop releases actuator drive cleanly.

Control, iteration 2

The system described above operates in joint space: the VLA commands the whole body joint by joint, and the physics layer realizes it. Iteration 2 moves the interface itself, from joint space to end-effector space, and is already in use.

 iteration 1   model → whole-body joint targets → physics layer → motion
 iteration 2   model → hand poses → the robot resolves the rest

In end-effector space the model commands where the hands go, and the robot resolves everything beneath, arms, base, and lift together. Models train on handheld-gripper human demonstrations, which broadens what counts as training data: demonstrations can be collected with a gripper alone, no robot present, and the same recordings remain valid as the platform beneath them evolves. Implementation details are deliberately withheld at this stage; quantitative results will accompany the release.

Training is the other half of this story: the technical report is coming soon.

← all posts · questions or corrections: hi@os3robotics.com