How ActDistill Slashes Deployment Costs of VLA Large Models

Vision‑Language‑Action (VLA) large models have become a hot research direction because they can perceive scenes, understand language commands, and directly output robot actions, effectively closing the perception‑to‑control loop. However, as model capabilities grow, deployment becomes increasingly burdensome: inference latency, compute demand, and action‑loop speed are highly sensitive in real‑world robot scenarios.

Why Existing Efficient VLA Methods Fall Short

Previous acceleration techniques—such as caching visual‑language features, layer pruning, dynamic skipping, or replacing heavy modules—primarily optimize the visual‑language processing pipeline. They share a blind spot: they do not explicitly target the transformation from visual‑language understanding to actionable robot motions. In VLA tasks, the model must produce continuous, executable actions, requiring not only image and instruction comprehension but also predictions of end‑effector trajectories, contact timing, and critical state transitions. Consequently, not all layers contribute equally to action generation; some layers are crucial for motion planning while others are less relevant.

ActDistill: Action‑Guided Self‑Derived Distillation

ActDistill addresses this gap by treating the full VLA model as a teacher and training a lightweight student to inherit only the action‑relevant capabilities. Its core idea is to separate computations that help the model understand the world from those that directly influence world interaction.

Graph‑Structured Encapsulation : Convert intermediate representations of the teacher into a graph that explicitly encodes dependencies relevant to actions.

Action‑Guided Distillation : Align the student not only on semantic outputs but also on the action‑oriented supervision derived from the teacher’s graph.

Dynamic Routing : At inference time, activate only the layers required for the current input and action demand, skipping redundant computation.

The method is named “Action‑Guided Self‑Derived Distillation” because the supervision signals are extracted from the teacher’s own structured knowledge ( self‑derived) and the distillation is driven by action relevance ( action‑guided).

Why a Graph Representation?

VLA representations intertwine spatial relations, visual semantics, language conditions, and action intents. A graph makes explicit which entities and relationships are critical for motion decisions—e.g., the relative position of a cup to a microwave, the geometry between a handle and a gripper, or the spatial constraints between target containers and obstacles. Unlike dense attention, which encodes these relations implicitly, the graph isolates the most influential dependencies for the student to learn.

Experimental Evaluation

ActDistill was evaluated on the LIBERO and SIMPLER embodied benchmarks, covering architectures such as OpenVLA and CogACT. Across multiple settings, the method reduced compute to roughly half of the original model while delivering up to 1.67× speedup. Success rates remained comparable, and in some long‑horizon or complex tasks, the lightweight model even showed modest gains.

Crucially, the paper reports real‑world tests on an ARX5 robot arm. The distilled model achieved the same average success rate as the full OpenVLA but cut average execution time from 10.2 s to 6.3 s (≈ 1.62× faster). Unlike cache‑only baselines, ActDistill maintained performance under real‑world noise and distribution shift, demonstrating deployment robustness.

Key Insights

Selective layer skipping based on action relevance yields “informed” compute reduction rather than blind pruning.

Dynamic routing enables the model to allocate more computation to challenging scenes (e.g., occlusions, fine contacts) while conserving resources on simple tasks.

Distilling action‑oriented knowledge can suppress irrelevant semantic noise, sometimes improving control quality.

Overall, ActDistill shows that efficient VLA deployment is achievable by focusing on the VL‑to‑Action chain, using graph‑structured knowledge extraction, action‑guided supervision, and runtime‑adaptive routing.