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

The paper introduces PoLAR (Polar Latent Actions with Radial structure), a novel approach for latent action pretraining that separates transition extent and mode by imposing a radial structure on latent actions. By utilizing temporal offsets between observations to inform the radius of latent actions, PoLAR enhances the representation of diverse transition modes, particularly in hyperbolic space, leading to improved policy performance in both simulation and real-world robot experiments. This advancement underscores the significance of latent action space geometry in effectively transferring visual pretraining to robot policy learning tasks.