RD-VLA Scales Robot Reasoning in Latent Space
Loading video
Loading videoRD-VLA uses a weight-tied recurrent action head to refine robot actions in latent space, avoiding explicit reasoning tokens while allocating test-time compute adaptively through convergence.
VLALatent ReasoningRecurrent NetworksTest-Time ComputeRobot ManipulationRD-VLAVisual-Language-Action
Category: research
Author: @DJiafei
Date: 2026-02-10T00:00:00
Duration: 82.9s
Reference: https://rd-vla.github.io/
