EgoAlign trains a humanoid robot using only first-person view videos of humans, without a single robot demonstration

The framework adapts movements and scale to the robot’s body, reconstructs its states in simulation, and then the deployed model directly moves objects and reaches unknown targets.

Filming humans in subjective view is much cheaper than teleoperating a robot, but these demonstrations are not suitable for training a humanoid: different body size, different controller responses, and no recorded robot condition.

EgoAlign closes the gap using the model and simulator of the target robot. During collection, a return of execution guides the human operator; Then, the locomotion steps are retained while the geometry of the upper body gestures is scaled and refined with the controller in the loop. A final causal rereading reconstructs the states of the robot and the motion labels corresponding to each observation.

Deployed without training

A vision-language-action model refined only on these adapted data is deployed as it is on a physical humanoid. It moves objects over long distances, joins new positions and makes interactions with the foot. Refinement improves the actual capture compared to a simple kinematic alignment, and human collection reduces the time spent on site compared to teleoperation.