PoLAR: Factorizing Extent and Mode in Latent Actions for Robot Policy Learning

Researchers at Seoul National University introduce PoLAR, or Polar Latent Actions with Radial structure, to separate transition extent from transition mode in robot latent actions. Instead of forcing one code to represent both how far a state changes and what kind of change occurs, PoLAR uses temporal offset as weak ordinal supervision: larger observation gaps are pushed to larger radii, while directions encode mode. Its latent space is implemented in hyperbolic geometry, where angular capacity expands with radius, and training combines start-conditioned ordering and origin-centered radial losses. For discrete actions, a factorized codebook represents each transition with one radial token and four direction tokens. In large-scale pretraining, the method uses frozen DINOv2 features, BridgeData V2’s 60,096 demonstrations, and a Prismatic-7B VLA. PoLAR reaches 70.2 percent average success with joint fine-tuning on RoboMimic and MimicGen, versus 37.0 percent for a flat tokenizer, and 62.5 percent on SimplerEnv-WidowX, exceeding UniVLA at 46.9 percent and π0.5 at 56.2 percent. On three real-world WidowX tasks, it achieves a 76.7 percent overall final-task average, outperforming SmolVLA, UniVLA, Villa-X, and π0.5. Diagnostics show that hyperbolic PoLAR’s latent actions explain 0.184 of ground-truth action variance, compared with 0.160 for its Euclidean variant, supporting the claim that latent-action geometry improves transfer from visual pretraining to robot control.