Unitree G1 humanoid catches itself on a ladder, a move absent from its human lessons, team says

In the climbing example, the G1 catches itself on a ladder. According to RoboHub, which reports what the team claims, this manoeuvre was not explicitly part of the human demonstrations used for training.

The account also says that the transition from one movement to another is not programmed by hand.

According to the post, the team trained its system, called ACP, on 13.2 minutes of human motion so that it learns where to find support. It then used reinforcement learning in simulation to adapt it to new terrain.

According to the published description, once the robot is moving, its onboard sensors and a goal to reach determine where it places its hands, feet and knees, with no reference motion to imitate.