Embodiment Alignment

Embodiment alignment is the training stage in which a robot policy that has learned general manipulation capabilities during pre-training is adapted so that its outputs are valid, safe, and effective on a particular robot's body, sensors, and control interface.

Mechanically, embodiment alignment usually combines a small amount of real-robot demonstration data with the pre-trained model's existing representations. Common approaches include supervised fine-tuning on embodiment-specific action trajectories, action-space translation layers that map the policy's generic action tokens onto the target robot's joint or end-effector space, and reinforcement-learning objectives that keep the policy within the kinematic envelope of the target body. The goal is to preserve the breadth of pre-training while replacing the abstract action semantics with the precise, embodiment-grounded commands the actual robot requires.

The benefit is that a single pre-trained backbone can be deployed across many robot families without retraining from scratch, dramatically reducing the per-robot data requirement. The cost is that misalignment can introduce safety risks: a policy that is fluent in general manipulation can still issue commands outside the reachable workspace, beyond torque limits, or at unsafe speeds for a specific body, and these failure modes are not always caught by standard pre-training evaluations. Embodiment alignment also competes for training capacity with instruction-following alignment, which is usually co-trained in the same stage.

It is unclear how little embodiment-specific data is genuinely sufficient — claims of useful alignment from a few hours of demonstrations are common but are rarely stress-tested on long-horizon, dexterous, or safety-critical tasks. The right balance between embodiment alignment, instruction alignment, and continued pre-training is still an open design choice. There is also no widely accepted benchmark that isolates embodiment-alignment quality from general capability, which makes cross-paper comparisons difficult.