Safety Evaluation of VLAs

VLA: Vision-Language-Action Model

Overview

This field tests whether robots and driving agents can interpret instructions and act safely in interactive environments. Benchmarks use matched safe and unsafe scenarios, trajectory-level safety rules, counterfactual tests, and simulation-to-physical evaluation to separate safe behavior from task completion. Methods include adversarial red teaming, runtime action screening and abstention, semantic guards, robustness checks, and safety-aware training. Strong hazard recognition or task success often fails to ensure safe execution, leaving open challenges in cross-domain generalization, standardized metrics, and balancing safety with useful performance.

Latest papers

41 papers14 papers in the last four weeks, up 250% on the four weeks before. 0.1% of all new papers.

Nov 15, 2026

ReactBench-AD: Do Driving VLAs and World-Action Models React to Imminent Hazards in Their Path?

Vision-Language-Action models (VLAs) and World-Action Models (WAMs) have emerged in autonomous driving with the promise of strong generalization to long-tail scenarios. Although these models achieve state-of-the-art results on autonomous driving benchmarks, most existing benchmarks are dominated by nominal scenarios and therefore cannot test this promise. To test one such long-tail capability that is critical for safety, we introduce ReactBench-AD, a benchmark that measures how driving models react to objects that appear in the ego vehicle's path. Built on real NAVSIM and PhysicalAI-Autonomous-Vehicles scenes, ReactBench-AD composites a crossing pedestrian, a standing pedestrian, or a stationary car consistently across all camera views and history frames. Each edited scene thus forms an exact, pixel-level counterfactual pair with its original, which allows us to isolate the causal effect of the inserted object on the predicted trajectory. Evaluating four VLAs and two WAMs, we find that they react weakly, often barely at all. Finally, we test how easily one of the evaluated VLAs, ReCogDrive, can be taught to brake hard for imminent hazards: a short fine-tuning stage on edited scenes substantially improves its reactions, with the learned behavior generalizing to unseen objects and fully real scenes, at minimal cost to nominal driving. These results call into question the out-of-the-box long-tail capabilities of these models, but suggest that they can be a strong foundation for acquiring them with little additional training, highlighting the value of real or synthetic data that demonstrates the long-tail behaviors nominal datasets rarely contain.
Nov 15, 2026

When Backdoors Can Be Switched Off: Suppression-Aware Backdoor Attacks on Vision-Language-Action Models

Vision-Language-Action (VLA) models map visual observations, robot states, and natural-language instructions to robotic actions, but their multimodal interfaces introduce new security risks. Existing VLA backdoor attacks primarily study trigger-based activation and stealthy action manipulation, leaving conditional suppression of malicious behavior underexplored. We investigate suppression-aware backdoors, which combine conjunctive trigger activation with instruction-conditioned suppression. We instantiate this design through three variants: Cross-Modal Trigger, combining robot-state and spatially separated visual triggers; Self-Suppressing Trigger, requiring a state trigger and a [DISRUPT] instruction token for activation, with [RESTORE] token suppressing the attack; and Multi-Trigger, additionally requiring a visual trigger while retaining suppression even when physical triggers persist. We evaluate these attacks using SmolVLA on two SO-100 manipulation datasets against four baselines: State Backdoor, BadVLA, TabVLA, and SilentDrift. Self-Suppressing and Multi-Trigger reduce Isolation Forest detection from 96--98% to 1--18% and action-space l divergence by 10--14X, achieving 63--71% attack success rates with perfect clean-task performance. Experiments with OpenVLA on LIBERO further show that the attacks transfer to a 7B-parameter policy, evading STRIP and action-distribution-shift detection. Our findings establish conditional suppression as an important dimension for evaluating backdoor threats in VLA policies.
Nov 6, 2026

Localizing and Editing the Goal Representation of a Vision-Language-Action Model

Operators can state safety constraints for robots as prohibitions such as "do not put the bowl on the stove". We study whether OpenVLA follows such prohibitions and where it represents the goal it executes. The 10 tasks of LIBEROGoal share one scene, so we can label the goal that each rollout executes. We combine closed-loop rollouts with activation patching, linear probes, activation steering and runtime monitors. On LIBERO-Goal, negation had no detectable effect on which task OpenVLA executed. The model executed the forbidden task in 79% of negated rollouts and completed 78% of plain tasks. A goal probe gives similar read-outs under "not X" and "X". Activation patching shows that the first action depends mostly on the instruction tokens before layer 12 and on the last prompt token from layer 16 onward. A fixed steering vector at the instruction tokens at layer 8 redirected 12 of 18 rollouts. Fixed vectors and prefill patches at layers 12 to 20 redirected none of 158. From layer 16, only patches at every decoded action token redirected the robot (15 of 18 rollouts at layer 24). When the instruction changes during a rollout, the layer-32 goal read-out detects non-compliance better than failure detectors, and an external probe on image and instruction features performs comparably. These results support enforcing prohibitions outside the instruction and monitoring goal read-outs against the specification. Goal edits in OpenVLA should act before layer 12 or patch every action token.
Nov 5, 2026

AffectiveVLA: Safe and Intent-Aware VLA Control for Human-Humanoid Social Interaction

Social touch interactions, such as handshakes, fist bumps, and hugs, require humanoid robots to infer human intent, generate contact-rich motions, and maintain safe physical contact. We present AffectiveVLA, an intent-gated Vision-Language-Action (VLA) pipeline deployed on a Unitree G1 humanoid with dexterous hands. A lightweight RGB-based model recognizes interaction intent from two onboard cameras and converts it into a language prompt for the VLA policy. Low-confidence predictions suppress motion execution, reducing unintended activations. During physical contact, interaction-specific compliance profiles adjust arm impedance, while an independent voice-triggered mechanism initiates a latched safe-stop procedure. Using VR teleoperation, we collected approximately 1,300 multimodal interaction episodes and fine-tuned four VLA models: GR00T, RLDX-1, Being-H0.5, and UnifoLM-VLA-0. On-robot evaluation identifies GR00T as the best-performing standalone model. Integrating intent recognition increases its macro-F1 score from 0.75 to 0.90, while the intent model achieves a p95 inference latency of 6.27 ms. These results demonstrate that explicit intent gating improves the reliability of VLA-based social interaction without compromising real-time operation.
Nov 3, 2026

Aero-VPA: Few-Step Flow-Matching Control over Non-Reconstructive JEPA Latents for Agile Flight on Edge Hardware

Vision-Language-Action (VLA) models generalize across tasks, but their autoregressive decoders need sequential forward passes in proportion to the action horizon, each bounded below on embedded aerial platforms by streaming billions of parameters from memory. We present Aero-VPA, a lightweight Vision-Predictor-Action architecture for 3D navigation on edge hardware. A compact RGB-D Vision Transformer and a five-head ensemble predictor, trained without pixel reconstruction, forecast the next latent state; a Conditional Flow Matching solver generates a velocity and yaw-rate chunk in two Euler steps, so each decision takes five sequential network evaluations; and the ensemble dispersion drives an uncertainty fallback. Version 1 specified the architecture, deployment and protocol without results. This version reports the first closed-loop evaluation: PX4 software-in-the-loop simulation of quadrotor missions in procedurally generated forests. Behaviour cloning failed on all 60 missions despite open-loop errors of 0.02-0.04 m/s: the policy copies its last command, a copycat form of causal confusion. DAgger with a privileged, safety-gated supervisor gave the first successes but plateaued at 11.1-21.1% across six checkpoints. A map-free runtime shield with a 15 s depth memory, barrier constraints, a cap on motion outside the 73-degree field of view and backoff along the flown path raised the first DAgger policy's success on 180 paired missions from 21.1% to 52.2% (95% CI [45.0, 59.4]; exact McNemar p = 1.2 x 10^-10) and cut its crashes from 72.2% to 22.2%, while timeouts rose from 2.8% to 23.3%. All six checkpoints gained 30.0-34.4 percentage points, at mean speeds of 1.16-1.36 m/s. The code (https://github.com/ylee-sling/Aero-VPA), the trained policy (72.6 MB, https://huggingface.co/ylee-sling/aero-vpa) and the full benchmark dataset are public.
Nov 1, 2026

When Delay Meets Drift: Physical Safety of Robot Foundation Models Is a Cross-Layer Problem

Robot foundation models now act in a closed loop, yet their safety is evaluated one variation at a time: robustness benchmarks perturb the scene while assuming instantaneous inference, and real-time execution methods study inference delay in nominal scenes. We argue that physical safety is not a property of the model alone, but emerges from how variations in the physical environment, the model, the computing platform, and execution interact in closed loop. We formalize this as cross-layer variation propagation across these system layers and use an interaction contrast to measure when joint variation creates more risk than the sum of its parts. In 4,000 paired closed-loop episodes of the pi0.5 vision-language-action model on LIBERO, a 400 ms inference delay raises the safety-violation rate by 3.5 percentage points under the training viewpoint but by 37 percentage points under a 60-degree camera shift: 33.5 percentage points (95% CI: 25-43) more than the two effects combined. We close with a research agenda toward predictable physical intelligence: predicting when variation becomes risk, and adapting compute, power, and control accordingly.
Oct 30, 2026

VLA Policies Do Not Know When Not to Act: Unintended Task Execution and Start Suppression

A robot should act only on an instruction from a person or on an intent that the robot explicitly represents and generates. Unintended actions taken without such an instruction can cause risk or inconvenience to people and the environment. We examine how recent vision-language-action (VLA) policies respond when the instruction is missing or meaningless. We apply a paired protocol with empty, neutral, nonsense, infeasible, and wrong-task instructions to six model checkpoints in LIBERO-Goal simulation. Holding the initial state fixed and varying only the instruction, we ask three questions: whether the robot starts to move, which task it completes, and what the instruction determines. Even when the instruction was empty, the robot started moving in every episode with all six checkpoints, and with five of them it often completed a task that the instruction had not specified. On two GR00T checkpoints, the instruction determined which task was executed but not whether the robot started moving; the robot moved even on infeasible instructions. We then tested three methods for suppressing the start: prepending "if no task is given do not move" to the instruction; a gate that holds the robot when the first action chunk for the given instruction nearly equals that for an empty instruction; and fine-tuning the policy to hold still on non-task inputs. The prefix stopped the robot in none of the six checkpoints, and the gate held it on non-task inputs in five of the six checkpoints. The fine-tuned policy never completed a task on non-task inputs, but it did not reliably keep the robot still, and it failed to suppress the start on infeasible instructions.
Oct 30, 2026

TRACE: Target-State Risk-Aligned Conditional Execution for Embodied Tracking

Embodied visual tracking requires a robot to follow a language-specified person under occlusion, look-alike distractors, and ambiguous instructions. We show that improving tracking persistence can simultaneously increase safety violations. Using a compact 0.6B vision-language-action (VLA) tracker with a frozen backbone, we update only the planner head with proximal policy optimization (PPO). PPO improves success and following rates across all three EVT-Bench settings, but increases the collision rate on distracted tracking because episodes that previously ended in target loss can instead end in contact. We introduce TRACE, an execution-time target-state risk gate that preserves the learned policy and intervenes selectively. A lightweight risk head processes 84 cues, including candidate boxes, identity margins, a learned range-and-bearing probe, and the policy's commands. When the calibrated risk exceeds a threshold, TRACE rescales the translational component of the next waypoint while leaving yaw unchanged. The executed action is identical to PPO's on more than 98% of steps. In paired evaluations on identical episodes, TRACE reduces the distracted-tracking collision rate by 3.96 percentage points, while further improving success and following rates. Matched slowdown controls without target-state timing do not reproduce this improvement. Concurrent evaluation reproduces a collision reduction of 3.77 points, compared with 0.29 points for identical-arm null pairs, indicating that the effect exceeds evaluation variability. These results show that target-state-aware execution can recover the safety cost of policy post-training with minimal intervention.
Oct 26, 2026

Beyond Behaviour: A Stimulus-Response Safety Evaluation of Vision-Language-Action Models

Humans exhibit biological sensitivity to safety-critical stimuli that can trigger defensive action. Across three models and two video benchmarks, we investigate whether similar responses emerge in vision-language-action models by examining their internal representations and action tendencies in response to visible hazards and harmful consequences. Our findings suggest that task execution may remain the default response even when danger is internally registered.
Oct 26, 2026

Benchmarking Obstacle-Conditioned Safety in Household Robot Navigation

Safe household navigation requires robots to regulate clearance and speed according to what an obstacle is and how close it is, without unnecessary conservatism. Yet existing evaluations do not test whether semantic safety distinctions are reflected in continuous motion. To that end, we introduce OSMA-Bench, an Obstacle-Conditioned Safe Motion Adaptation Benchmark built on spatially matched scenario sets that differ only in obstacle type. We also propose the Semantic Sensitivity Index (SSI), which measures whether clearance and distance-dependent speed and smoothness follow the caution ordering of obstacles. Evaluating VLM-based and vision-language-action (VLA) policies on the benchmark, we find that neither consistently adapts its motion to obstacle safety, even with explicit safety prompts. Our analysis shows that the policies can correctly distinguish obstacles by safety relevance but fail to reflect this distinction in the motion parameters or actions they produce. For VLM-based policies, which recognize the obstacle and then generate a navigation program, adding an explicit parameter-inference step between these two stages improves adaptation. These findings expose a gap in semantic-to-motion grounding, highlighting the need to translate semantic caution into appropriately safe motion.
Oct 23, 2026

When Human Corrections Travel: Safety in Federated VLA Adaptation

Vision-language-action (VLA) policies can adapt to local human intervention, while federated learning can combine distributed robot experience without pooling raw records. We examine their prospective intersection: sharing adaptations derived from diverse human corrections across tasks, environments, people, and robot embodiments. Corrections differ in modality, safety semantics(urgency, behavioral scope, etc.), and privacy sensitivity. A shared update that improves mean task success may suppress a rare safety intervention or produce an unsafe response on another robot. We argue for three requirements: preserve the context and uncertainty of multimodal corrections when deciding what to share; assess candidate updates in each receiving robot's closed loop before deployment; and evaluate leakage alongside retention of safety-critical evidence. Immediate, legible defensive behavior remains necessary during human interaction. We offer counterexamples and a falsifiable evaluation agenda, not experimental results.
Oct 22, 2026

Progress Probes P2P^2: Lightweight Runtime Monitoring for Vision-Language-Action Models

Vision-Language-Action (VLA) models can perform diverse manipulation tasks directly from visual observations and natural language instructions, but they remain prone to mistakes and unpredictable failures. Deploying these policies without continuous human supervision will require robust runtime monitors that detect when the policy has gone off-course. Existing monitors use external foundation models, such as vision-language models (VLMs) to supervise the policy, introducing significant complexity and computational cost for monitoring at runtime. In this work, we investigate an alternate approach: can a VLA's own internal activations provide the information needed to monitor its execution? Our method trains linear probes on policy activations from demonstrations to estimate progress, then flags executions that deviate from expected progress. In experiments with the open-source VLA π0.5π_{0.5} and across simulated and real-world manipulation tasks, this approach is competitive with or outperforms VLM-based monitors in task failure detection, as well as methods which require failure data for training. These findings suggest that simple probes of VLA internals can be used to develop reliable runtime monitors, opening a path toward building policies that can act reliably and monitor their execution without external foundation models.
Oct 20, 2026

Perceives but Doesn't Brake: A Perception-Action Binding Gap in Reasoning Driving VLAs

Reasoning-augmented vision-language-action (VLA) driving models can perceive and reason about safety-critical hazards yet fail to translate that evidence into safe actions. We characterize this failure as a perception-action binding gap: hazard information is available inside the model but exerts insufficient control over trajectory generation. Using matched simulated counterfactuals and logged driving replay, we show that in-path hazards are actually encoded in the visual representation, but longitudinal plans are dominated by ego-speed history and respond only weakly to hazard proximity or the model's own reasoning trace. Further analysis shows that the trajectory generator is capable of braking but sometimes fails to do so due to the perception-action binding gap. Without retraining or modifying any weights, we inject a hazard-conditioned energy into the frozen flow-matching sampler of Alpamayo-1.5, restoring the missing coupling at generation time. With exact obstacle geometry, guidance cuts footprint intrusion from 58.1% to 14.5% and recovers safe braking whenever sufficient stopping distance remains, eliminating collisions entirely in our braking-relevant closed-loop battery. It also reverses the diagnosed dependence of the plan, from tracking ego speed to responding to obstacle range. These results separate hazard perception from hazard control in driving VLAs, show that unsafe behavior can arise even when the model has both the evidence and the ability to act on it, and propose a simple fix for bridging this binding gap.
Oct 19, 2026

Refuses When Asked, Complies When Acting: A Runtime Safety Check for Robot Control Loops

Frontier models can control robots through agent harnesses that loop on one question: what motion comes next toward the user's goal? We show that this default loop carries out harmful instructions. With GPT-6 Astra controlling a low-cost SO-101 arm through the Inspect Robots harness on three household tasks whose hazard is common knowledge, such as "put the knife in the toy basket", the policy refused 0% of 15 trials, attempted the harm every time, and completed it in 27%. Its one exit, give_up, is meant for goals that cannot be achieved, and it never cited the hazard. We add an optional runtime safety check to the loop: before the first motion, the model is asked whether it would carry out the instruction, given the same camera frames. The check would have stopped all 15 hazardous instructions, including the 4 that completed the harm, and flagged none of the 15 harmless controls. Code and data: https://github.com/Arnav-Gr0ver/Refuses-When-Asked-Complies-When-Acting.
Oct 14, 2026

SafeDriveVLA: Navigation-Conditioned World Model Dreaming for Conflict-Aware End-to-End Autonomous Driving

Vision-Language-Action (VLA) models are gaining attention in autonomous driving, inheriting commonsense priors from large language models (LLMs) and widening the navigation interface from a fixed command set to free-form natural-language instructions. However, this open-ended interface exposes a new vulnerability surface: an inattentive driver may issue instructions that are unsafe given the current scene, and a malicious user may issue dangerous ones on purpose. We present the first benchmark suite that evaluates navigation following and conflict awareness together, along with adversarial robustness. Existing VLAs fall short on both: SimLingo completes no language-instructed lane change in closed loop, and all models degrade sharply under conflicting signals. We propose SafeDriveVLA, the first driving VLA that (1) explicitly predicts a three-level navigation-compliance mode to decouple conflict reasoning from action generation, and (2) uses navigation-conditioned world model dreaming to inform this prediction. Unlike prior world models that roll out expert or planned actions, ours dreams of strictly following the navigation signal, showing when following is safe. SafeDriveVLA reaches 88.16 Driving Score (DS) and 68.64% Success Rate (SR) on Bench2Drive, lifts closed-loop language following by 49.3% and reduces collisions under unsafe instructions from 60 to 38 compared to the best SimLingo variant. Video demonstrations are available on our project page.
Oct 5, 2026

What the Guard Misses, the Robot Executes: Implied Harm in VLA Instructions

Vision-language-action models (VLAs) act on instructions without being able to refuse, so screening harmful requests falls to monitors. We test whether these monitors catch ordinary robot tasks requested for harmful reasons, holding the task fixed while varying only how explicitly the intent is stated. π0.5π_{0.5} completes the task at every level of explicitness, as often as for harmless controls. Text guards flag nearly every blunt request but few implied ones: up to 95% of implied-harm runs end with the task done and no flag raised, and most still leak after recalibrating on robot instructions at low false-positive rates. Monitoring the model's activations does not close this gap. Linear probes separate harmful from harmless instructions almost perfectly in the base language and vision-language models, but this separation weakens after robot training in two model families, most for implied harm.
Oct 2, 2026

ManiPhysicsBench: Physics-Based Assessment of Object Preservation in VLA Manipulation

Vision-language-action (VLA) models aim to perform diverse manipulation tasks, but task success in existing rigid-body benchmarks does not indicate whether they preserve objects. We introduce ManiPhysicsZoo, which consolidates literature-supported material properties, 3D meshes, and supporting references into reusable object assets. Using these assets, a solver-based assessment computes grasp-specific damage thresholds from object geometry, material properties, and recorded grasp conditions and compares them with recorded contact forces to assess potential deformation and fracture. Building on these components, ManiPhysicsBench evaluates object preservation in LIBERO and SimplerEnv across three physics axes and three difficulty levels. Public VLA checkpoints show a substantial gap between task success and safe success, defined as task completion while preserving the object. Their gripper commands concentrate near full opening and closure, with largely similar aggregate distributions across objects, consistent with binary gripper supervision. We examine how object-specific continuous gripper labels change model behavior by retraining a VLA model. The retrained model shows more object-dependent gripping and higher safe success, but lower task success and limited generalization of object-preserving behavior.
Sep 26, 2026

Are Vision-Language-Action Models Robust to One-Step Observation Perturbations?

Understanding the safety risks of vision-language-action (VLA) models is essential for their deployment in the physical world. Existing safety research has mainly considered persistent perturbations that are applied continuously to observations throughout an episode. However, momentary observation corruption, in which observations are severely perturbed only briefly within an episode, remains an underexplored safety threat. To address this gap, this work investigates robustness to one-step perturbations applied at a single time step per episode. Our experiments reveal that these perturbations substantially degrade VLA performance and that their impact depends on the action chunk execution length. Based on them, we propose CARE, which dynamically selects the execution length based on consistency with the previously predicted action chunk. CARE improves robustness with low computational overhead while preserving clean performance by selecting shorter execution lengths only under perturbations.
Sep 20, 2026

ReVeal: A Reconstruction-Aware Real-to-Sim Framework for VLA Policy Evaluation

Simulation-based evaluation provides a scalable and repeatable alternative to real-world evaluation of vision-language-action (VLA) policies. However, reconstruction errors can cause simulated policy performance to diverge from real-world performance, motivating the need to assess reconstructed environments for downstream VLA policy evaluation. We present ReVeal, a real-to-sim assessment framework combining workspace reconstruction, reconstruction-level assessment, and matched closed-loop policy evaluation. Novel-View Mesh Fidelity (NVMF) and Annotated Planar Geometry Fidelity (APGF) assess observation and planar geometric fidelity, respectively. We also develop PGSR-D, a reconstruction pipeline incorporating monocular depth supervision to improve geometry where multi-view visual cues are limited. Across 8 assessment scenes, NVMF and APGF consistently distinguish the fidelity of 2DGS, PGSR, and PGSR-D. Matched evaluations of GR00T, SmolVLA, and pi0.5 across 8 humanoid manipulation tasks show consistent ordering between reconstruction fidelity and real-sim performance agreement across pipelines. Further analysis of the evaluation workspaces shows that higher fidelity is associated with stronger real-sim agreement.
Sep 18, 2026

Beyond the Leaderboard: Counterfactual Diagnosis of End-to-End and VLA Driving Policies Under Domain Shift

End-to-end and vision-language-action (VLA) driving policies are compared by leaderboard rank, but a rank reports an outcome, not the behaviour behind it, so it predicts poorly how a policy will behave at a new site. On six released policies, rank on nuScenes open-loop error or on NAVSIM's leaderboard does not carry over to scenes with a pedestrian near the ego corridor at a new site. We propose a counterfactual check-up: a few hundred real frames, each edited two ways (pedestrian removed, or re-lit by a night-style perturbation), every edit verified by an independent detector, and the change in the planned trajectory read as a diagnosis rather than a score. From these edits two causal axes are read, and five exams built on them separate what a score merges: how far the policy plans to drive, whether seeing the pedestrian buys safety, whether that response scales with danger, whether the plan moves when nothing requires it, and how much an irrelevant lighting change moves it. On 246 NAVSIM near-pedestrian scenes, in the cells where the pedestrian lies on the planned path only 1.9% of responses are genuine avoidance, and under our open-loop protocol the median clearance change is at most 0.03 m and the median change in planned distance at most 0.08 m for every policy. In a pre-registered test from left- to right-hand drive, the exposure and specificity orderings, the lighting verdict and the collision outcome transfer, while point values and the hazard-sensitivity verdict do not. Read as a selection report, the profiles say which policy is safe because it plans short, which covers a human-like distance without yielding, and which is unsteady under a change that requires no reaction, and they price each verdict: most settle within a few dozen frames, hazard sensitivity needs hundreds. Code and edited frames will be released.
Sep 18, 2026

SafeStage: Evaluating Safety Before, During, and After Vision-Language-Conditioned Robot Manipulation

Vision-language-conditioned robot policies integrate perception, language understanding, and control for general-purpose manipulation. However, existing evaluations often focus on task success, isolated physical constraints, semantic refusal, or realized physical damage, providing limited insight into where safety fails during closed-loop manipulation. We introduce SafeStage, a lifecycle-structured benchmark for evaluating manipulation safety before, during, and after task execution. SafeStage contains 97 purpose-built risk scenarios organized into three stages. Initial-State Hazards captures safety-relevant relations that must be resolved before manipulating the target. Execution-Time Safety evaluates unsafe contacts, trajectories, region entries, and object interactions during execution. Final-State Hazards capture unstable or otherwise unsafe conditions remaining after nominal task completion. The benchmark evaluates realized interactions using event-based and state-based checks and reports native task success independently from stage-specific safety outcomes. We evaluate representative direct-action Vision-Language-Action (VLA) policies and policies with world-model-based policies under a common closed-loop protocol. Our results demonstrate that nominal task completion frequently coexists with safety violations and that different policies exhibit distinct failure profiles across the three stages. By separating task success from safety and localizing when violations occur, SafeStage provides a unified diagnostic testbed for evaluating and improving vision-language-conditioned robot manipulation policies.
Sep 11, 2026

STAGE: Diagnosing Semantic Transfer at Grounded Execution in Embodied Agents

Embodied language grounding requires more than identifying the referent of an instruction: recovered semantics must also control the action an agent exposes. We study this missing link as a semantic-action gap, where instruction semantics are recoverable but weakly expressed in native continuous actions. We introduce SAT-Bench, a fixed-observation counterfactual benchmark that holds the visual scene and agent state fixed while changing only instruction semantics. On LIBERO target-name and pixel-grounded relation swaps, target recovery reaches 100.0% and 95.8%, whereas OpenVLA action sensitivity remains only 6.8% and 7.7%. The gap persists across 1,000 additional compositional and temporal/procedural counterfactuals, with overall action sensitivity of 6.1%. Hidden-state, threshold-free, cross-policy, and rollout diagnostics further support this semantic-action transfer failure. We introduce VISA, a lightweight execution-time interface that converts recovered semantics into ALLOW, DEFER, target-consistency, and verified-selection decisions. VISA reduces invalid-instruction blind execution from 92.7% to 2.8% while preserving 94.0% of normal commands, and verified selection further improves target-consistent action exposure without updating the underlying policy. Overall, embodied language evaluation should measure semantic-action transfer, not semantic parsing alone.
Aug 18, 2026

LIBERO-VIFO: Benchmarking the Capability and Safety of Visual Cue Following in Vision-Language-Action Models

Visual cues are increasingly adopted to guide robot learning, but whether Vision-Language-Action (VLA) models can reliably follow authorized cues while disregarding unauthorized ones remains unclear. Existing work covers only a narrow range of cue forms and focuses on final task success, providing only a coarse assessment of cue-following capability. Treating all visual cues as authorized also leaves safety risks of unauthorized following unexplored. To address these gaps, we introduce LIBERO-VIFO, a benchmark to evaluate both the capability and safety of visual cue following in VLA models. LIBERO-VIFO defines eight visual cue families spanning diverse forms. A total of four protocols in two parts are defined: Part I tests cue understanding and authorized following, while Part II evaluates unauthorized visual cue following under language-cue conflict and empty language conditions. Evaluating seven VLA models reveals that although visual cue understanding does not reliably translate into execution, current VLAs are able to execute cue-indicated tasks without language instruction, exposing an emerging risk of unauthorized visual cue following. Extended experiments on scene-instantiated cues, safety-critical settings, and real-robot deployment corroborate these findings. LIBERO-VIFO brings both the capability and safety of visual cue following into systematic evaluation, establishing visual-centric safety as a new perspective for the VLA community.
Aug 18, 2026

MANIGUARD: A Benchmark and Data Suite for Specification-Grounded Safety Evaluation and Improvement of Robotic Manipulation

Foundation-model policies for robotic manipulation are advancing rapidly on task success, but rigorous evaluation of whether they succeed safely is still lacking. We introduce ManiGuard, a specification-grounded framework for evaluating and improving the safety of foundation-model manipulation, comprising the ManiGuard-Bench task suite and a paired safety-annotated trajectory-generation pipeline. ManiGuard-Bench organizes six contact-rich household task families into 200 locked base tasks along a skill ×\times constraint taxonomy, with safety specified independently of task success. Each task is evaluated under one in-distribution and four single-axis out-of-distribution perturbations that hold the safety specification fixed, giving 1,000 locked scenarios. Every rollout is runtime-checked by LTLf_f-grounded automaton monitors over physics-grounded predicates rather than learned classifiers or LLM judges, in simulation and on a physical Franka platform. The pipeline pairs an automated motion-planning generator with human teleoperation, annotated by the same per-step monitor, and directly supports safety-aware fine-tuning; we release 8,000 safety-annotated demonstrations, 40 per base task. Benchmarking zero-shot and fine-tuned VLAs across more than 23,000 rollouts, we find: (i) safety must be evaluated independently of task success, as 6-21% of successful rollouts violate the specification; (ii) fine-tuning on our suite raises safe task completion from near zero to 7.5-29.8% and engaged-and-safe behavior from 16-40% to 51-72%; but (iii) a gap remains that scaling demonstrations does not close, with 21-42% of engaged rollouts still violating, two of six families below 2% safe success for every policy, and these failures persisting under distribution shift and on hardware.
Aug 17, 2026

MAE I Trust Myself? Self-Evaluating VLA Action Generation with Markov Attention Entropy

Vision-Language-Action models (VLAs) integrate visual perception, language instruction, and action generation into end-to-end policies across heterogeneous architectures. However, enabling VLAs to self-evaluate their action generation reliability without external supervision remains a major challenge. Existing methods either rely on expert annotations or estimate uncertainty only from output statistics, largely ignoring internal signals. In this work, we observe that internal visual modality entropy exhibits consistent distinctions between successful and failed tasks across heterogeneous VLAs. Although VLAs' architectures differ in their action generation, we show that they share a common latent action generation abstraction evolving under visual perception, language instruction, and State Input, which we formulate as a Conditional Generative Markov Chain. Based on this formulation, we propose MAE (Markov Attention Entropy), a self-evaluation framework that directly converts internal attention signals into architecture-aware reliability scores, and introduce LIBERO-Reflect, a 4,000-episode benchmark combining 2,000 standard episodes and 2,000 challenging episodes across four subsets. Extensive experiments across heterogeneous VLA architectures and diverse scenarios show that MAE consistently outperforms state-of-the-art baselines on AUPR, AUROC, and FPR@95.
Aug 17, 2026

SafeGesture: Evaluating Fine-Grained Hand Gesture Understanding in Vision-Language Models through Scenario-Conditioned Safety Interpretation

Open-weight and frontier vision-language models (VLMs) perform well on general image understanding, but their ability to interpret fine-grained hand gestures in safety-critical operational contexts remains largely unexamined. We introduce SafeGesture, a benchmark that evaluates whether a model can infer scenario-appropriate safety actions from hand gestures. It pairs six HaGRID gestures with eight operational scenarios for 4,800 items and evaluates Qwen2.5-VL-7B, LLaVA-NeXT-7B, InternVL2-8B, Phi-3.5-Vision, and GPT-4o. Results reveal a perception-reasoning decoupling: GPT-4o achieves 98.4% gesture accuracy but 53.3% safety accuracy, while Qwen2.5-VL reaches 84.9% and 39.5%, yielding gaps of 45.0 and 45.4 percentage points. Four of five models rarely or never use the uncertainty label, and failure directions differ substantially across models. Accuracy also obscures label bias: a scenario-majority policy with no visual input reaches 58.3%, above every evaluated model, while only GPT-4o exceeds this prior under macro-F1. Visual input improves safety accuracy by 11.2 to 30.2 percentage points, but providing the ground-truth gesture as text improves performance by only 0.4 to 3.2 points, and no model exceeds 56.2%. These results indicate that the main bottleneck is scenario-conditioned safety reasoning rather than gesture recognition.
Aug 14, 2026

SSP: An Event-Matched Syn2Sim2Phy Cross-Domain Evaluation Framework for Autonomous Driving VLA Models

Vision-language-action (VLA) models for autonomous driving jointly produce scene interpretation, language-based reasoning, and driving trajectories. Existing evaluations often use independently selected synthetic, simulated, and physical data, so measured performance gaps can be confounded by changes in scenario content rather than genuine domain sensitivity. We propose SSP (Synthetic-Simulation-Physical), an event-matched Syn2Sim2Phy evaluation framework that anchors cross-domain comparison to the same safety-critical interaction. Starting from a synthetic long-tail video, SSP builds a validated event specification that preserves road topology, participant roles, relative motion, conflict evolution, passing order, response constraints, and event phases. Platform-specific realizations are then constructed in CARLA and on a closed proving ground and are evaluated only after transfer audits confirm preservation of mandatory event properties. SSP maps heterogeneous outputs from OpenEMMA, LLaViDA, and Alpamayo-R1 into common semantic slots and a 1 s trajectory window to assess output validity, semantic accuracy, critical-interaction recognition, trajectory quality, and risk response. Across Cut-in and vulnerable-road-user crossing cases, the macro-averaged Integrated VLA Capability Scores are 0.259, 0.291, and 0.325 in the Synthetic, Simulation, and Physical domains, respectively, while the best domain varies by scenario. Alpamayo-R1, OpenEMMA, and LLaViDA obtain scores of 0.405, 0.338, and 0.131. SSP provides a reproducible scene-transfer chain and an evidence-qualified evaluation of VLA behavior without assuming that the Physical domain is universally superior.
Jul 20, 2026

Reasoning as a Double-Edged Sword: Architecture and Cross-Stage Robustness in Vision-Language-Action Models

Does adding a reasoning step make a Vision-Language-Action (VLA) model more robust to perturbation? Intuitively, a policy that reasons before acting should absorb a perturbed input better than one that maps observations directly to actions. We test this premise head-on across three models that span the reasoning spectrum (no reasoning, a text chain-of-thought, and a latent iterative loop), perturbing each at the vision, reasoning, and action stages on LIBERO and SimplerEnv. Two questions organize the study: does the reasoning design shift robustness, and can the reasoning be read back at runtime as a safety signal? We find that the latent-iterative model is by far the least robust: under both stochastic noise and white-box perturbation its task success collapses, while the other two hold. This fragility is structural rather than cumulative: varying the reasoning depth at inference barely moves it. Reasoning outputs can in principle be monitored, but the monitors fail under fair tests. A plan--action consistency probe that looks near-perfect under naive evaluation falls to chance under adaptive attack. Under matched-FPR calibration, fusing it with an action-anomaly probe never lifts defended success above undefended. Scoped to these output-level behavioral probes under white-box vision-stage attack, this ceiling is a precondition that any viable defense must first satisfy.
Jun 25, 2026

ForesightSafety-VLA: A Unified Diagnostic Safety Benchmark for Vision-Language-Action Models

In embodied intelligence, safety is a prerequisite for reliable robot deployment in the physical world. Current vision-language-action (VLA) models continue to advance toward general-purpose task capability, yet their embodied safety limits remain poorly understood. To address this gap, we introduce ForesightSafety-VLA, a diagnostic benchmark that makes safety the primary evaluation target for VLA systems. We define a 13-category safety taxonomy covering physical interaction safety (Safe-Core), instruction-side safety (Safe-Lang), and perception-side safety (Safe-Vis), and evaluate policies under three controlled dimensions of variation -- scene structure, language command, and visual observation -- so that failure sources can be diagnosed rather than hidden in a single aggregate score. Beyond binary task success, ForesightSafety-VLA measures process-level risk through cumulative safety cost (CC) and risk exposure time (RET), together with a four-quadrant decomposition of safe/unsafe success and failure. We instantiate 66 safety-augmented base scenarios in RoboTwin across 5 embodiments and report results on representative VLA baselines. Across the evaluated baselines, even the strongest policy incurs non-trivial safety cost and unsafe nominal success, while structure and visual variation induce substantially stronger safety degradation than ordinary language variation. These results suggest that embodied safety is tightly coupled to perception, grounding, and control competence rather than being reducible to post-hoc safety filtering alone.
Jun 22, 2026

LIBERO-Safety: A Comprehensive Benchmark for Physical and Semantic Safety in Vision-Language-Action Models

Despite the impressive manipulation capabilities of Vision-Language-Action (VLA) models, their operational safety under strict constraints remains largely unverified. To address this, we introduce a parametric safety benchmark to procedurally generate safety-critical scenarios with comprehensive stochasticity. To overcome the scalability bottlenecks of human teleoperation, we develop a novel keypose-driven data generation pipeline. Leveraging this infrastructure, we curate a large-scale dataset of 19,664 strictly collision-free demonstrations with extensive domain randomization. We then conduct a systematic cross-paradigm evaluation of eight VLA and two embodied foundation models. Our analysis reveals a critical generalization-safety tension: although high-diversity training fosters safer trajectories, task success remains fundamentally bottlenecked by sub-optimal trajectory synthesis and semantic misalignment. By providing a scalable pipeline, a robust dataset, and profound failure-mode insights, LIBERO-Safety establishes a crucial foundation for developing safe and reliable VLA models.