Description:
The open-loop trajectories simulate a rollout of 1 second with fixed action sequence starting from an initial state. The video is displayed at 0.25x. The bright robot corresponds to the rollout using the true model. The dark robots simulate the action sequence using the learned model. Each dark robot corresponds to an ensemble component. For each model type, the open-loop rollouts of the best and worst seed are shown. For all systems except the humanoid, the learned models produce a plausible rollout, which might not be accurate but is visually feasible. Only the humanoid generates divergent and obviously false trajectories.