Robot policy gains are coming from action interfaces and execution constraints
Robot policy work this day is concentrated on executable action design. Vision-Language-Action (VLA) papers tune the action head, latent action alignment, task adapters, and onboard latency.