Industry Monitor Humanoid Industrial & Cobot AGV / AMR Quadruped Reducers · Servos · Sensors Drones & Autonomy Embodied AI
Robos News
Robotics

Freeze, Share, Shrink: Rethinking the Action Backbone in Diffusion Policies

arXiv:2511.12101v3 Announce Type: replace Abstract: Many recent Vision-Language-Action models employ diffusion or flow-matching backbones with hundreds of millions of parameters for action generation. However, unlike image synthesis where the output spans millions of diverse pixels, a manipulation policy generates only short sequences of low-dimensional, physically correlated action values, a far simpler target that may not require such capacity. We confirm this intuition and show that, in modu

Published September 15, 2026 · Category: Robotics

Overview

arXiv:2511.12101v3 Announce Type: replace Abstract: Many recent Vision-Language-Action models employ diffusion or flow-matching backbones with hundreds of millions of parameters for action generation. However, unlike image synthesis where the output spans millions of diverse pixels, a manipulation policy generates only short sequences of low-dimensional, physically correlated action values, a far simpler target that may not require such capacity. We confirm this intuition and show that, in modulation-conditioned diffusion policies, task adaptation can be routed entirely through the conditioning pathway, leaving a frozen, observation-free backbone that serves as a reusable trajectory prior. To establish this, we introduce a decoupled training recipe: a general-purpose action head is first pretrained on observation-free forward-kinematics data, then frozen while only the conditioning pathway is trained for downstream tasks. Using Diffusion Policy as a testbed, we show that on both MimicGen and LIBERO, a single frozen backbone shared across all tasks matches normally trained counterparts. In our ablations, this succeeds with modulation-based conditioning, while attention-based conditioning is embedded in the backbone's own weights and collapses once they are frozen. Ablations show the pretraining signal (joint positions, end-effector poses, or none) has little effect, while a randomly initialized backbone fails entirely, indicating that pretraining need only supply a general trajectory prior. Finally, a 5M-parameter MLP backbone matches or exceeds both the 244M U-Net and the transformer on these benchmarks, suggesting the action backbone is over-parameterized and that architectures inherited from image and language generation are a poor inductive-bias fit for the low-dimensional action target, with implications for action-backbone design in VLA models and other policies with an action backbone.

Source

Originally published at arxiv.org.

Related Articles

Robos News Newsroom

Robos News reports on robotics research, components, manufacturers, field deployments, and industrial automation worldwide. Tip our newsroom: [email protected]

Email the newsroom →
Reporting standard: Product specifications, deployment counts, and performance claims are attributed to their source. Safety-critical decisions should be based on the applicable technical documentation and validation for the operating environment.
More from News →