A direct-action policy predicts each action chunk independently, so the motion jumps where chunks meet. ODE policy replaces only the action head: it learns how the action changes over the robot's ...
Python 3.11+, NumPy and SciPy; no GPU or deep-learning framework is required. The current implementation is a research release. The scalar/group/column Gaussian routes use exact spectral calculations.
Some results have been hidden because they may be inaccessible to you
Show inaccessible results