A new arXiv paper introduces ShieldVLA, a method for safety alignment in vision-language-action models used in robotic manipulation and navigation. The authors argue that existing fine-tuning approaches, which largely depend on Lagrangian optimization, offer only limited safety guarantees. The work instead frames safety as a feasibility-aware alignment problem.
A new arXiv paper argues that the action-generation backbones in Vision-Language-Action models are far larger than the task requires, since robot actions carry much less information than image pixels. The authors introduce a freeze-share-shrink strategy to cut parameters in diffusion and flow-matching policies while preserving performance. The work targets more efficient manipulation models for robotics.
A new arXiv preprint proposes ReWeight, a technique for post-training vision-language-action models when in-domain robot demonstrations are scarce. The approach retrieves relevant egocentric human demonstrations and applies sample weighting to make better use of that human data, since collecting robot-specific data is expensive. The paper targets adapting VLA models to particular robots and tasks.
A new arXiv paper introduces IMPACT-VLA, a method for tracing how much each input modality — camera images, proprioceptive state, and language instructions — contributes to a vision-language-action policy's decisions at different points during task execution. The approach relies on counterfactual trajectories to isolate the effect of individual inputs, addressing the difficulty of interpreting these multimodal robot policies. The abstract excerpt does not detail experimental results or benchmarks.
A new arXiv paper proposes TIDAL, a control scheme that alternates between diffusion-based planning and action execution to keep vision-language-action models running at high frequency. The authors argue that current VLA systems rely on a low-frequency batch-and-execute approach, and that the resulting mismatch between model inference speed and robot control rate creates gaps in which the agent cannot react. TIDAL aims to close that blind spot while retaining the semantic generalization of large VLA models.
Researchers released vla-eval, an evaluation harness designed to simplify how vision-language-action models are tested across multiple simulation benchmarks. The tool addresses the friction of conflicting dependencies and inconsistent evaluation protocols that arise when benchmarks are combined in a single pipeline. It aims to make VLA evaluation more reproducible and easier to extend with new benchmarks.
A new arXiv paper introduces ActSafeGuard, a differentiable approach for enforcing hard physical constraints in vision-language-action and world-action models used for robotic manipulation. The method aligns constraint enforcement with training so that generated actions remain feasible and safe. The authors argue that current models can produce actions that violate physical limits, making them unsafe for deployment.
A new arXiv paper proposes a time-frequency geometric cross-attention mechanism for vision-language-action policies that emit chunks of actions in one forward pass. The authors argue that an action chunk is effectively a short multivariate trajectory and design their architecture to model it as such. The work targets robotic control models that generate one to two seconds of coordinated motion per prediction.
Researchers report the first bit-flip attack against a vision-language-action model, showing that Rowhammer-style faults can corrupt INT8 weights in deployed quantized systems. Using only a few gradient-selected bit flips, they drive the policy's closed-loop success rate down to near zero. The paper finds that a VLA model's action-decoding architecture largely determines how exposed it is to such faults.
A new arXiv paper introduces VLA-Precision, a method for fine-tuning pretrained vision-language-action models with online reinforcement learning directly on real robots. It targets manipulation tasks where such models still struggle, particularly those requiring precise, repeatable motions. The proposed asymmetric co-bootstrapping approach aims to make real-world trial-and-error learning more efficient and autonomous.