Summary
首个 VLA 安全领域系统性 Survey。用 attack timing (training/inference) × defense timing 的双 timing 轴组织威胁-防御配对,覆盖 data poisoning、backdoor、对抗扰动、semantic jailbreak、物理攻击等威胁面,以及 training-time 对齐、runtime monitoring、physical fail-safe 等防御,最后系统梳理 benchmark/metric、六大部署场景与 future directions,并配套持续更新的 Awesome-VLA-Safety repo。
Problem & Motivation
VLA 正成为具身智能的统一基座,但其安全挑战与 text-only LLM 有本质区别:(1) embodied consequences——物理动作部署后不可逆;(2) multimodal attack surface——视觉观测、语言、proprioceptive state 都可被利用;(3) real-time 约束——latency-heavy 防御在毫秒级控制中可能失效;(4) error compounding——单点扰动在 long-horizon 序列中级联放大;(5) training pipeline 风险——未审查的 demonstration 数据引入供应链漏洞。现有文献分散在 robot learning、adversarial ML、AI alignment、autonomous systems safety 等社区,缺乏统一视角。
Method
Survey 的核心组织框架是 2×2 双 timing 轴:attack timing (training/inference) × defense timing (training/inference)(Fig 2),用于原则化地配对威胁与缓解,并暴露覆盖空白。各 section:
- Sec 2 Background:VLA 形式化为 POMDP(视觉/proprioceptive 观测、离散/连续/chunked 动作、语言条件、BC 目标);架构组件(visual encoder CLIP/SigLIP、language backbone LLaMA/Gemma、action decoder token-based/diffusion/flow-matching);训练范式(VL 预训练、robot demo 微调、preference alignment、LoRA);代表系统 RT-1/2、Octo、OpenVLA、π₀/π₀.₅、SpatialVLA。
- Sec 3 Training-time Attacks:input-centric backdoor(token/pixel trigger、cross-modal composite trigger、物理 object trigger);temporal/state-space backdoor(利用 action-chunk 盲区的平滑时序扰动、sequential error accumulation、proprioceptive state poisoning)。
- Sec 4 Training-time Defenses:data/perception/reward-centric alignment(stage-aware 监督、self-evolving + pose-based exploration、thermal/depth 多模态感知增强);policy-centric safety optimization(constrained MDP、post-training safety unlearning、online rejection sampling);human-in-the-loop(从干预中做 action preference alignment)。
- Sec 5 Inference-time Safety:semantic jailbreak(prompt injection 利用 output-action mismatch、white-box adversarial suffix);visual perturbation(cross-modal mismatch 优化、action freezing、physical semantic deception);physical intervention(object placement 误导导航、sensor signal injection);防御侧 decision-layer guardrail、runtime monitoring、physical fail-safe。
- Sec 6 Evaluation:benchmark(adversarial robustness、task-level safety、comprehensive capability-safety、jailbreak/alignment、runtime monitoring)+ metric。
- Sec 7 Deployment:六大场景 + cross-domain challenge。
- Sec 8 Future Directions:certified robustness、physically realizable defenses、safety-aware training、unified runtime architecture、standardized evaluation、lifecycle safety。
Key Results
Survey 类,贡献是框架性与 taxonomic:
具名威胁/防御(精读后补全):
- Training-time attack:BadVLA(objective-decoupled 干扰)、DropVLA(cross-modal alignment 劫持)、GoBA(3D 物理 object trigger,黑盒)、SilentDrift(smootherstep 时序扰动钻 action-chunk 盲区)、Clean-Action(sequential error trap)、State Backdoor(遗传算法做 proprioceptive 投毒);
- Training-time defense:SafeVLA(constrained MDP 联合优化 task-safety)、SORL(safety critic 多目标)、EvoVLA(stage-aligned reward + pose-based exploration + 长时记忆)、APO(从干预做 action preference alignment)、Hi-ORS(outcome-based rejection sampling)、Safe-Night VLA(thermal/depth 增强 + CBF-QP runtime filter)、VLA-Forget(post-training safety unlearning);
- Inference-time attack:RoboPAIR(黑盒 semantic jailbreak)、BadRobot(output-action mismatch)、Adv-Robo(white-box discrete prompt 优化)、VLA-Fool(cross-modal mismatch 操纵轨迹)、FreezeVLA(视觉扰动致 action freezing 瘫痪)、AARONS(物理 object shift 误导导航)、Phantom Menace(直接 sensor signal injection)。
Metric 体系:task-level(SVR Safety Violation Rate、RejR Rejection Rate、SR)、behavioral(CR Collision Rate、SS Safety Score、SPL)、robustness(ASR Attack Success Rate、PDR Performance Drop Rate、Certified Robustness Radius)、composite(safety-performance trade-off、cost-aware evaluation)。
关键观察:training-time 漏洞(poisoning/backdoor)相对 inference-time 攻击仍 underexplored;action chunking 制造的时序盲区被攻击者主动利用;多数工作的 safety 机制 sim-to-real 迁移未验证;尚无 comprehensive VLA safety 的标准 benchmark。
Strengths & Weaknesses
亮点:
- 首次系统整理 VLA 安全领域、填补空白,双 timing 轴框架清晰,把威胁与防御按缓解阶段配对便于定位 mitigation;
- 覆盖广:从训练数据投毒到 runtime fail-safe、从 benchmark 到六大部署场景;具名工作整理较全(攻击/防御各 7+ 个代表系统),可作 reading list;
- 配套 Awesome-VLA-Safety repo 持续更新、有 community contribution 机制。
局限:
- Survey 类工作,无新方法/实验,仅文献整理;价值取决于 taxonomy 是否真有组织力,双 timing 轴本质是直觉性的 2×2,对”哪类威胁最危险/最易防”缺乏量化优先级;
- 部分威胁(freezing attack、proprioceptive state-space poisoning)描述简略,需回溯原始论文;
- 未深入 safety-capability trade-off 的具体量化方法,Pareto frontier 只作为 future direction 提出。
潜在影响:为 VLA safety 提供统一坐标系,催化 robot learning ↔ adversarial ML ↔ AI alignment 跨社区对话,指出 certified robustness for embodied trajectories、physically realizable defenses 等关键 open problem,可能成为后续研究热点。
Mind Map
mindmap root((VLASafety)) Problem 物理后果不可逆 多模态攻击面 实时约束 vs safety干预 长时域误差累积 数据供应链漏洞 Method 双timing轴 2x2 框架 Training-time Attack BadVLA DropVLA GoBA SilentDrift state poisoning Training-time Defense SafeVLA SORL EvoVLA APO Hi-ORS VLA-Forget Inference-time Attack RoboPAIR BadRobot FreezeVLA Phantom Menace Inference-time Defense decision-layer guardrail runtime monitoring physical fail-safe Results Benchmark taxonomy Metrics SVR RejR ASR PDR 6 deployment domains Future directions
Notes
- 与 GUI Agent 安全有交叉:多模态攻击面、实时约束、runtime monitoring 均相关,可借鉴其双 timing 轴组织 GUI agent 的威胁-防御。
- Certified robustness for embodied trajectories 是关键 open problem——VLA trajectory 需要不同于 image/text 的 robustness 证明(per-step ε-ball 不够,要 trajectory-level 保证)。
- Physically realizable defenses 的难点在 sub-100ms latency 约束下做实时防御,这是 VLA 区别于 LLM safety 的硬约束。
- EvoVLA(self-evolving + pose-based exploration)可与 SpatialEvo 的 self-evolving 思路对照——同样 self-evolve,但一个 verifier 是几何 oracle,一个是 safety reward。
- action chunking 的时序盲区是 VLA 特有攻击面,值得单独追踪(SilentDrift/Clean-Action)。