Act with Intent: Distilling Behavior Intent for Vision-Language-Action Models

The development of Vision-Language-Action (VLA) models has been significantly accelerated by pre-trained Vision-Language Models (VLMs). However, most existing end-to-end VLAs treat the VLM primarily as a multimodal encoder, directly mapping vision-language features to low-level actions. This paradigm underutilizes the VLM's potential in high-level decision making and introduces training instability, frequently degrading its rich semantic representations. To address these limitations, we introduce DIAL, a framework bridging high-level decision making and low-level motor execution through a differentiable latent intent bottleneck. Specifically, a VLM-based System-2 performs latent world modeling by synthesizing latent visual foresight within the VLM's native feature space; this foresight explicitly encodes intent and serves as the structural bottleneck. A lightweight System-1 policy then decodes this predicted intent together with the current observation into precise robot actions via latent inverse dynamics. To ensure optimization stability, we employ a two-stage training paradigm: a decoupled warmup phase where System-2 learns to predict latent futures while System-1 learns motor control under ground-truth future guidance within a unified feature space, followed by seamless end-to-end joint optimization. This enables action-aware gradients to refine the VLM backbone in a controlled manner, preserving pre-trained knowledge. Extensive experiments on the RoboCasa GR1 Tabletop benchmark show that DIAL establishes a new state-of-the-art, achieving superior performance with 10x fewer demonstrations than prior methods. Furthermore, by leveraging heterogeneous human demonstrations, DIAL learns physically grounded manipulation priors and exhibits robust zero-shot generalization to unseen objects and novel configurations during real-world deployment on a humanoid robot.

A fine-tuned Vision-Language-Action (VLA) policy can easily place the alphabet soup in the basket on command, just as it saw during fine-tuning. However, if the basket is shifted five centimeters to the left, it drops the soup off the table. We developed an interpretability pipeline to investigate what such a policy is computing. Injecting task A's activations into the action expert while the policy looks at task B's scene makes pi0.5 execute task A's motor trajectory in 99.6% of episodes ($n{=}1{,}968$). The injected program is bound to absolute workspace coordinates and ignores the visible scene, which provides a mechanistic account of the perturbation brittleness that concurrent benchmarks report. The binding holds on a physical UR10e: directed replay redirects the commanded task in 26/26 rollouts in one direction and 18/26 in the other, with no-injection and random-direction controls at the floor, and once the source object is moved 10--15,cm the arm grasps and releases at the recorded coordinates in 50/50 trials, picking up the wrong block when one sits there and closing on empty air when none does. Applying the same interventions to six VLAs on LIBERO, MetaWorld, SimplerEnv, and ALOHA over 420{,}000+ rollouts yields three further findings: every architecture encodes the instruction, and four of the five weight it below vision whenever the scene alone identifies the goal; under the standard replacement write-back the SAE pooling that preserves behavior is architecture-specific (pi0.5 per-token, X-VLA mean-pool, SmolVLA indifferent), and goal selection and motor generation separate into VLM and expert pathways wherever the two are distinguishable. We release Action Atlas (https://action-atlas.com), an interactive platform covering all 520+ trained SAEs, 79 identified concepts, and the scored UR10e rollouts. Code: https://github.com/CWRU-AISM/action-atlas

Billion-parameter Vision-Language-Action (VLA) policies have recently shown impressive performance in robotic manipulation, yet their size and inference cost remain major obstacles for real-time closed-loop control. We introduce VLA-AD, a distillation framework that uses a Vision-Language Model as an offline semantic supervisor to transfer large VLA teachers into lightweight student policies. Instead of relying only on low-level action imitation, VLA-AD augments teacher-provided 7-DoF action targets with high-level semantic guidance, including task phase anchors and multi-frame operating-direction descriptions. These auxiliary signals are used only during training: at test time, the student policy runs independently, with neither the VLA teacher nor the VLM required. We evaluate VLA-AD on three LIBERO benchmark suites. Using OpenVLA-7B as the teacher, our method produces a 158M-parameter student, yielding a $44\times$ reduction in model size while matching the teacher with only a $0.27\%$ average relative gap. The resulting policy runs at 12.5 Hz on an RTX 4090, achieving a $3.28\times$ inference speedup over OpenVLA-7B. We further show that the same semantic distillation pipeline generalizes to a different $π_{0.5}$-4B teacher, where the student outperforms the teacher on two suites and remains within $0.53\%$ on libero\_goal. Additional analysis indicates that phase-level supervision and multi-frame directional cues make the student less sensitive to noisy teacher actions, such as erroneous high-frequency gripper changes. Overall, VLA-AD demonstrates that offline semantic guidance from VLMs can substantially improve the efficiency, robustness, and deployability of VLA policy distillation.