V-Link: Recovering Lost Visual Representations in Action DiT for Vision-Language-Action Models

Vision-Language-Action (VLA) models have recently emerged as a promising paradigm for robotic manipulation, in which reliable action prediction critically depends on accurately interpreting and integrating visual observations conditioned on language instructions. Although recent works have sought to enhance the visual capabilities of VLA models, most approaches treat the LLM backbone as a black box, providing limited insight into how visual information is grounded into action generation. Therefore, we perform a systematic analysis of multiple VLA models across different action-generation paradigms and observe that sensitivity to visual tokens progressively decreases in deeper layers during action generation. Motivated by this observation, we propose DeepVision-VLA, built on a Vision-Language Mixture-of-Transformers (VL-MoT) framework. This framework enables shared attention between the vision foundation model and the VLA backbone, injecting multi-level visual features from the vision expert into deeper layers of the VLA backbone to enhance visual representations for precise and complex manipulation. In addition, we introduce Action-Guided Visual Pruning (AGVP), which leverages shallow-layer attention to prune irrelevant visual tokens while preserving task-relevant ones, reinforcing critical visual cues for manipulation with minimal computational overhead. DeepVision-VLA outperforms prior state-of-the-art methods by 9.0% and 7.5% on simulated and real-world tasks, respectively, providing new insights for the design of visually enhanced VLA models.

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

Vision-Language-Action (VLA) models have shown strong potential for robotic manipulation, but many existing methods still rely on direct mappings from language and visual observations to dense actions. This formulation can weaken the semantic reasoning capability inherited from pre-trained Vision-Language Models (VLMs), which are mainly optimized for visual-linguistic understanding rather than low-level control, and becomes fragile under spatial variations, including changes in object positions, scene layouts, robot embodiments, and camera viewpoints. To address these limitations, we propose H-VLA, a hierarchical VLA framework that decouples high-level key-action reasoning from low-level motion generation. H-VLA combines a Key-Action Model for predicting a key-action as the next manipulation subgoal, a Motion Planning Model for generating dense future actions conditioned on the predicted key-action, and a Unified Camera-Centric Action Space for consistent representation across datasets, embodiments, and viewpoints. We further adopt a two-stage training strategy that emphasizes key-action reasoning during pre-training and dense motion generation during fine-tuning. Experiments show that H-VLA achieves strong performance on SimplerEnv, reaching 91% on Google Robot visual matching, 84% on Google Robot variant aggregation, and 81% on WidowX visual matching. On Agilex real-robot tasks, H-VLA improves over the strongest baseline by 10, 47, and 16 percentage points under in-distribution, out-of-distribution position, and out-of-distribution scene/object settings, respectively.