mlsanityData to 7 Oct 2026

SafeVLA: Towards Safety Alignment of Vision-Language-Action Model via Constrained Learning

arXiv 2503.03480 ↗First posted 5 Mar 2025cs.RO cs.AIcited by 69

On the map: Vision-language-action models → Safety and adversarial attacks

also in 2 more
Abstract (arXiv)

Vision-language-action models (VLAs) show potential as generalist robot policies. However, these models pose extreme safety challenges during real-world deployment, including the risk of harm to the environment, the robot itself, and humans. How can safety constraints be explicitly integrated into VLAs? We address this by exploring an integrated safety approach (ISA), systematically modeling safety requirements, then actively eliciting diverse unsafe behaviors, effectively constraining VLA policies via safe reinforcement learning, and rigorously assuring their safety through targeted evaluations. Leveraging the constrained Markov decision process (CMDP) paradigm, ISA optimizes VLAs from a min-max perspective against elicited safety risks. Thus, policies aligned through this comprehensive approach achieve the following key features: (I) effective safety-performance trade-offs, reducing the cumulative cost of safety violations by 83.58% compared to the state-of-the-art method, while also maintaining task success rate (+3.85%). (II) strong safety assurance, with the ability to mitigate long-tail risks and handle extreme failure scenarios. (III) robust generalization of learned safety behaviors to various out-of-distribution perturbations. The effectiveness is evaluated on long-horizon mobile manipulation tasks. Our data, models and newly proposed benchmark environment are available at https://pku-safevla.github.io.

Key facts. Cited by 69 papers in the map (data through 7 Oct 2026). Placed in Safety and adversarial attacks, one of 13 directions in Vision-language-action models. Builds on 64 of its 96 references that are in the map. First posted on arXiv 5 Mar 2025.

Builds on

64 of 96 references.

References · 64

  1. 2022Training language models to follow instructions with human feedbackcited by 15,712
  2. 2017Proximal Policy Optimization Algorithmscited by 14,441
  3. 2024The Llama 3 Herd of Modelscited by 14,192
  4. 2023Llama 2: Open Foundation and Fine-Tuned Chat Modelscited by 12,865
  5. 2023RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Controlcited by 2,993
  6. 2022Constitutional AI: Harmlessness from AI Feedbackcited by 2,950
  7. 2024OpenVLA: An Open-Source Vision-Language-Action Modelcited by 2,777
  8. 2024π0: A Vision-Language-Action Flow Model for General Robot Controlcited by 2,297
  9. 2022RT-1: Robotics Transformer for Real-World Control at Scalecited by 2,007
  10. 2022Objaverse: A Universe of Annotated 3D Objectscited by 1,522
  11. 2024Octo: An Open-Source Generalist Robot Policycited by 1,427
  12. 2023Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversationscited by 1,124
38 more
  1. 2017Vision-and-Language Navigation: Interpreting visually-grounded navigation instructions in real environmentscited by 1,013
  2. 2017AI2-THOR: An Interactive 3D Environment for Visual AIcited by 850
  3. 2023BeaverTails: Towards Improved Safety Alignment of LLM via a Human-Preference Datasetcited by 828
  4. 2022A Generalist Agentcited by 802
  5. 2022Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learnedcited by 702
  6. 2024RDT-1B: a Diffusion Foundation Model for Bimanual Manipulationcited by 670
  7. 2019RLBench: The Robot Learning Benchmark & Learning Environmentcited by 669
  8. 2025FAST: Efficient Action Tokenization for Vision-Language-Action Modelscited by 643
  9. 2023Safe RLHF: Safe Reinforcement Learning from Human Feedbackcited by 630
  10. 2021CALVIN: A Benchmark for Language-Conditioned Policy Learning for Long-Horizon Robot Manipulation Taskscited by 564
  11. 2023Reinforced Self-Training (ReST) for Language Modelingcited by 486
  12. 2023FoundationPose: Unified 6D Pose Estimation and Tracking of Novel Objectscited by 411
  13. 2022ProcTHOR: Large-Scale Embodied AI Using Procedural Generationcited by 369
  14. 2024Robotic Control via Embodied Chain-of-Thought Reasoningcited by 340
  15. 2021Safe Learning in Robotics: From Learning-Based Control to Safe Reinforcement Learningcited by 337
  16. 2020Responsive Safety in Reinforcement Learning by PID Lagrangian Methodscited by 302
  17. 2024TraceVLA: Visual Trace Prompting Enhances Spatial-Temporal Awareness for Generalist Robotic Policiescited by 298
  18. 2024A Survey on Vision-Language-Action Models for Embodied AIcited by 274
  19. 2023AI Alignment: A Comprehensive Surveycited by 256
  20. 2023Challenges and Applications of Large Language Modelscited by 251
  21. 2018Safe Exploration in Continuous Action Spacescited by 237
  22. 2025Hi Robot: Open-Ended Instruction Following with Hierarchical Vision-Language-Action Modelscited by 216
  23. 2023An Overview of Catastrophic AI Riskscited by 214
  24. 2024RT-H: Action Hierarchies Using Languagecited by 204
  25. 2021Simple but Effective: CLIP Embeddings for Embodied AIcited by 191
  26. 2020RoboTHOR: An Open Simulation-to-Real Embodied AI Platformcited by 190
  27. 2024Aligning Cyber Space with Physical World: A Comprehensive Survey on Embodied AIcited by 175
  28. 2024VLABench: A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Taskscited by 171
  29. 2022A Review of Safe Reinforcement Learning: Methods, Theory and Applicationscited by 169
  30. 2020Recovery RL: Safe Reinforcement Learning with Learned Recovery Zonescited by 156
  31. 2023Safety-Gymnasium: A Unified Safe Reinforcement Learning Benchmarkcited by 135
  32. 2023RT-Trajectory: Robotic Task Generalization via Hindsight Trajectory Sketchescited by 131
  33. 2024Llama Guard 3 Vision: Safeguarding Human-AI Image Understanding Conversationscited by 118
  34. 2018Learning-based Model Predictive Control for Safe Explorationcited by 116
  35. 2017AI Safety Gridworldscited by 108
  36. 2024GRAPE: Generalizing Robot Policy via Preference Alignmentcited by 106
  37. 2024Aligner: Efficient Alignment by Learning to Correctcited by 96
  38. 2025DexGraspVLA: A Vision-Language-Action Framework Towards General Dexterous Graspingcited by 87

50 most cited of 64 All 64 (JSON)

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2025: 12; 2026: 57.

69 papers.

Sort:

By area:

  1. 2025A Survey on Vision-Language-Action Models: An Action Tokenization Perspectivecited by 96
  2. 2025Pure Vision Language Action (VLA) Models: A Comprehensive Surveycited by 46
  3. 2025VLSA: Vision-Language-Action Models with Plug-and-Play Safety Constraint Layercited by 32
  4. 2025VLA-Arena: An Open-Source Framework for Benchmarking Vision-Language-Action Modelscited by 28
  5. 2025Safe RLHF-V: Safe Reinforcement Learning from Multi-modal Human Feedbackcited by 27
  6. 2025Optimizing Safe and Aligned Language Generation: A Multi-Objective GRPO Approachcited by 24
  7. 2025AVA-VLA: Improving Vision-Language-Action models with Active Visual Attentioncited by 22
  8. 2025FreezeVLA: Action-Freezing Attacks against Vision-Language-Action Modelscited by 19
  9. 2026Uncertainty Quantification in LLM Agents: Foundations, Emerging Challenges, and Opportunitiescited by 15
  10. 2026iWorld-Bench: A Benchmark for Interactive World Models with a Unified Action Generation Frameworkcited by 14
  11. 2025Generative RLHF-V: Learning Principles from Multi-modal Human Preferencecited by 13
  12. 2025EMMA: Generalizing Real-World Robot Manipulation via Generative Visual Transfercited by 12
38 more
  1. 2026Restoring Linguistic Grounding in VLA Models via Train-Free Attention Recalibrationcited by 11
  2. 2026SafeVLA-Bench: A Benchmark for the Success-Safety Gap in Vision-Language-Action Modelscited by 8
  3. 2026FlowPRO: Reward-Free Reinforced Fine-Tuning of Flow-Matching VLAs via Proximalized Preference Optimizationcited by 8
  4. 2026LIBERO-Safety: A Comprehensive Benchmark for Physical and Semantic Safety in Vision-Language-Action Modelscited by 8
  5. 2025CoReVLA: A Dual-Stage End-to-End Autonomous Driving Framework for Long-Tail Scenarios via Collect-and-Refinecited by 7
  6. 2026Faithful GRPO: Improving Visual Spatial Reasoning in Multimodal Language Models via Constrained Policy Optimizationcited by 6
  7. 2026HazardArena: Evaluating Semantic Safety in Vision-Language-Action Modelscited by 6
  8. 2026Safe-Night VLA: Seeing the Unseen via Thermal-Perceptive Vision-Language-Action Models for Safety-Critical Manipulationcited by 5
  9. 2026Escaping the Diversity Trap in Robotic Manipulation via Anchor-Centric Adaptationcited by 4
  10. 2026Learning What to Say to Your VLA: Mostly Harmless Vision Language Action Model Steeringcited by 4
  11. 2026UniIntervene: Agentic Intervention for Efficient Real-World Reinforcement Learningcited by 4
  12. 2026Neuro-Symbolic Safety Guidance for Vision-Language-Action Models via Constrained Flow Matchingcited by 4
  13. 2026LABSHIELD: A Multimodal Benchmark for Safety-Critical Reasoning and Planning in Scientific Laboratoriescited by 3
  14. 2026HomeGuard: VLM-based Embodied Safeguard for Identifying Contextual Risk in Household Taskcited by 3
  15. 2026VLA-Hijack: A Transferable Patch Attack against Vision-Language-Action Models via Visual Proprioception Hijackingcited by 3
  16. 2026Your Model Already Knows: Attention-Guided Safety Filter for Vision-Language-Action Modelscited by 3
  17. 2026Constrained Group Relative Policy Optimizationcited by 2
  18. 2026FeaXDrive: Feasibility-aware Trajectory-Centric Diffusion Planning for End-to-End Autonomous Drivingcited by 2
  19. 2026Towards Backdoor-Based Ownership Verification for Vision-Language-Action Modelscited by 2
  20. 2026World Pilot: Steering Vision-Language-Action Models with World-Action Priorscited by 2
  21. 2026SafeDojo: Safe Reinforcement Learning for VLA via Interactive World Modelcited by 2
  22. 2026Any-Body Guard: Universal Safeguarding for Manipulation Policies via Action Maskingcited by 2
  23. 2026LabGuard: Grounding Natural-Language Laboratory Rules into Runtime Guards for Embodied Laboratory Agentscited by 2
  24. 2026SoftVTBench: A Safety-Aware Visuo-Tactile Benchmark for Physically Constrained Robotic Manipulation of Deformable Objects (Early Version)cited by 2
  25. 2026Security of World-Model-Based Embodied AI: A Lifecycle of Threats, Defenses, and Evaluationcited by 2
  26. 2026ActFovea: Runtime Safeguarding for VLA Policies via Spatiotemporal Visual-Action Consistencycited by 2
  27. 2026Safe Vision Language Action Models via Barrier Enhanced Flow Matchingcited by 2
  28. 2025AERMANI-VLM: Structured Prompting and Reasoning for Aerial Manipulation with Vision Language Modelscited by 1
  29. 2026HomeSafe-Bench: Evaluating Vision-Language Models on Unsafe Action Detection for Embodied Agents in Household Scenarioscited by 1
  30. 2026RLearner-LLM: Balancing Logical Grounding and Fluency in Large Language Models via Hybrid Direct Preference Optimizationcited by 1
  31. 2026Uncovering and Mitigating Positional Blind Spots in Vision-Language-Action Modelscited by 1
  32. 2026CounterAlign: Counterfactual Supervision for Vision-Language-Action Modelscited by 1
  33. 2026Vision--Language Signals in Constrained RL: Safety Gains Without Anticipationcited by 1
  34. 2026OOD-MMSafe: Advancing MLLM Safety from Harmful Intent to Hidden Consequencescited by 0
  35. 2026Specification-Aware Distribution Shaping for Robotics Foundation Modelscited by 0
  36. 2026RedVLA: Physical Red Teaming for Vision-Language-Action Modelscited by 0
  37. 2026How VLAs Fail Differently: Black-Box Action Monitoring Reveals Architecture-Specific Failure Signaturescited by 0
  38. 2026VLESA: Vision-Language Embodied Safety Agent for Human Activity Monitoringcited by 0

50 most cited of 69 All 69 (JSON)

Similar, not cited

Close in wording, no citation either way

  1. 2026Towards General Language-Conditioned Latent Safety Filters
  2. 2026Vision-Language-Action Safety: Threats, Challenges, Evaluations, and Mechanisms
  3. 2026ForesightSafety-VLA: A Unified Diagnostic Safety Benchmark for Vision-Language-Action Models
  4. 2025SAFE: Multitask Failure Detection for Vision-Language-Action Models
  5. 2026SafeGen-LLM: Enhancing Safety Generalization in Task Planning for Robotic Systems
  6. 2024FOSP: Fine-tuning Offline Safe Policy through World Models
  7. 2026SafeLoop: Risk-Aware Rollback for Vision-Language-Action Manipulation
  8. 2025RobustVLA: Robustness-Aware Reinforcement Post-Training for Vision-Language-Action Models