DriftQL generates actions in one forward pass using a learned drift field. Attraction keeps candidates near the data, repulsion spreads them out, and the critic tilts the objective toward higher value. No ODE solver, no distillation network.
A single world model trained across eight quadrupeds that generalizes zero-shot to new robot morphologies by explicitly conditioning on physical specifications rather than baking in a fixed embodiment.
CDPs add a simple contraction regularizer to diffusion policies, pulling nearby sampling trajectories together to suppress solver and score-matching errors. This yields more robust action generation in offline RL and imitation learning, especially in low-data regimes.
VOCALoco predicts the viability and cost of transport for several pretrained locomotion skills from local heightmaps, then executes the safest, most efficient one. It improves robustness on stairs and transfers zero-shot to real hardware.
Imitation learning policies often fail when robots drift out-of-sample. We propose SCDS, a framework that models policies as contractive dynamical systems, ensuring all rollouts converge regardless of perturbations. The architecture guarantees contractivity by construction, enabling unconstrained optimization. We also provide formal bounds on worst-case and expected loss, and demonstrate strong out-of-sample recovery on manipulation and navigation tasks.