CopyQuad skips the depth-estimation stage that video-to-robot pipelines usually need. For side-view footage of a walking or trotting animal, it pins the one axis the camera can’t see to the robot’s own stance width, then measures what that costs against real depth models and motion capture.
Regularised RL has long promised robustness guarantees, but weakly regularised algorithms like PPO still dominate sim-to-real pipelines. We derive SMM-AC, a continuous-control actor-critic built on the soft mellowmax value function, and show on a Franka arm and an ANYmal-D quadruped that regularisation translates into real sim-to-real gains.
QWM conditions one world model on scale-invariant physical features and trains every policy inside it. Given the same morphology spec, a model-free policy collapses on unseen robots while QWM transfers zero-shot to new quadrupeds, in simulation and on hardware.