SymmGrid Framework Accelerates On-Robot Learning via Symmetry Augmentation

Researchers have introduced SymmGrid, a trajectory-level augmentation framework that leverages parallelized symmetries to accelerate on-robot deep reinforcement learning. The method models a Markov Decision Process under a symmetry tree, generating admissible invariant transformations that form a geometric grid structure. State representations use egocentric or exocentric images combined with proprioception, requiring homographies to warp visual scenes according to spatial transformations. These parallelized transformations produce diverse symmetric equivalences, populating the replay buffer with varied experiences to reduce wall-clock training times. The work addresses the bottleneck of slow training in physical robots, enhancing both egocentric and exocentric visual setups.