Reducing the cost of on-robot learning
SymmGrid, submitted July 29, is a trajectory-augmentation method designed to reduce wall-clock reinforcement learning time on physical robots. It generates admissible symmetry transformations of state-action pairs and aligns egocentric or external camera images with robot proprioception to expand the replay buffer.
Results on real manipulation
The team trained real robots on peg insertion, cable routing, and object relocation. Compared with the reported state of the art, convergence was 1.37–2.17 times faster and evaluation success was 1.09–1.27 times higher. Fastest convergence times were 16.6, 10.9, and 79.3 minutes, with trajectory-wide performance improvements up to 2.59 times.
The core idea is to reuse one experience as several when spatial transformations preserve the physical meaning of the action.
Meaning and limitations
Shorter on-site learning could reduce the cost of adapting robots to new layouts, but safety boundaries, collision monitoring, and human stop controls remain separate requirements.
These are preprint results from three manipulation tasks and specific hardware. Not every symmetry is valid in the real world, and the same gains are not established for deformable objects, mobile robots, or long-horizon work.