Researchers led by Lequn Fu have built ReLoco, which trains a single policy that lets a humanoid robot walk, get up from any fallen posture and go back to following its commands, according to a paper on arXiv. Learning all of this in one policy is hard because walking upright and recovering on the ground need different kinds of guidance.
ReLoco uses adversarial motion priors conditioned on the robot's posture. The posture picks which set of expert motions guides learning, and a shared discriminator gives separate outputs for the two behaviours. A stage in simulation supplies reference motions that the robot can actually perform.
In the authors' tests the full policy got up on its own 96.8% of the time and recovered from pushes 96.96% of the time, resuming its commands in both cases. On the 64.5 kg Tiangong 3.0 humanoid it walked, ran, got up and moved on after falls with one controller and no fine-tuning on the robot.



