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

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.