PAIR Builds a Bridge from Perception to Action in VLA Models

Introduction

Vision-language-action models face a more demanding problem than ordinary vision-language systems. They must interpret an observation and an instruction, then turn that interpretation into a continuous sequence of robot controls. In many existing continuous-action VLAs, the transition between perception and action is left implicit and is supervised mainly through the final action-prediction loss. PAIR addresses this gap by learning an explicit intermediate representation between the two spaces.

How PAIR works

The method is designed around a shared perception-action representation rather than a simple additional action head. During training, a Masked Action Autoencoder encodes chunks of expert actions into horizon-aligned Action Latent Tokens. This gives the action sequence a structured latent form that can be aligned with perception features, instead of treating it only as a set of continuous values to regress at the end.

A Bridge Module then extracts task-relevant features from the current visual-language representations. These features are aligned with the Action Latent Tokens to produce Bridge Tokens. The resulting representation is intended to retain information about the scene and instruction while also capturing the organization of expert behavior. The Bridge Tokens are projected into the action-token space and injected into the initial Action Tokens, giving the Action Expert an action-aware starting point before it refines the trajectory.

At inference time, the action autoencoder is removed. Bridge Tokens are generated only from the current observation and instruction, so the deployment pipeline does not require expert actions that would be unavailable in a real task.

Reported results

The paper evaluates PAIR on LIBERO, LIBERO-Plus, and CALVIN ABC-D, using evaluated versions of OpenVLA-OFT and VLA-Adapter. The supplied summary reports several results:

• On LIBERO-Plus, VLA-Adapter’s success rate increases from 59.1% to 64.2%.
• On CALVIN, its average completed sequence length rises from 4.42 to 4.53.
• The paper also reports gains across seven real-world tasks, although the provided material does not include the complete per-task numbers or experimental details.

Why it matters

PAIR highlights a potentially important design issue in VLA systems: the challenge may not lie only in understanding images and language or in decoding actions, but also in the interface between those stages. A bridge representation that carries both task semantics and action structure could provide the action module with more useful information before trajectory refinement begins.

The available material does not establish which component contributes most to the improvement. More complete ablations would be needed to separate the effects of the action latent space, feature alignment, and early token injection. The excerpt also does not provide enough information to assess computational overhead, scaling behavior, or generalization beyond the reported benchmarks.

Even with those open questions, PAIR offers a clear and reusable direction: instead of relying on an implicit jump from perception to control, VLA models can learn an explicit interface that organizes task understanding into action-ready representations.

Source: Hugging Face Daily Papers