Research map / Vision-language and multimodal / Robot action and navigation
SafeVLA: Towards Safety Alignment of Vision-Language-Action Model via Constrained Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes an integrated safety approach (ISA) that models safety predicates, elicits unsafe behaviors in procedurally generated scenes, and aligns VLAs (SPOC-based) with a CMDP Lagrangian PPO objective, evaluated on the new Safety-CHORES benchmark.
Problem
VLAs deployed in the physical world pose safety risks to environment, robot and humans, but safety constraints are not explicitly integrated into their training.
Approach
Frame VLA safety alignment as a CMDP solved by Lagrangian min-max RL (PPO-style, with separate reward and cost advantages), using cost signals from safety predicates over risk-eliciting procedurally generated mobile manipulation scenes.
Key results
- ISA reduces cumulative cost by 83.58% versus FLaRe while raising average success rate by 3.85%.
- On extreme failure scenarios, FLaRe incurs average CC 71.68 versus 2.20 for ISA.
- On DivScene zero-shot, ISA attains average SR 0.39 and CC 1.0, versus FLaRe SR 0.37 and CC 10.5.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add diverse risk elicitation using safety-critical components and rich scenes (ablation) | ISA · Safety-ObjNav | CC: 5.01 vs. 1.854 | Helped |
| add diverse risk elicitation using safety-critical components and rich scenes (ablation) | ISA · Safety-ObjNav | SR: 0.645 vs. 0.865 | Helped |
| use dynamic Lagrangian multipliers instead of fixed penalty coefficients (ablation) | ISA · Safety-ObjNav | SR | Helped |
| tighten the cost threshold to 10% of FLaRe's converged cost (ablation) | ISA · Safety-ObjNav | SR | Hurt |
| tighten the cost threshold (ablation) | ISA · Safety-ObjNav | CC | Helped |
Surprising findings
- Reward-shaping FLaRe-RS performed worse than the ablated ISA trained without safety-critical scenes on safety, indicating heuristic penalties are insufficient.
- ISA's cost is decoupled from task success, with even failed episodes remaining safe, unlike FLaRe where cost correlates negatively with success.
Limitations
- Binary cost signals ignore severity weighting of safety violations.
- Trajectory-level predicate credit assignment is simplistic (cost on final step of violating segment).
- Real-world validation is limited to a single Safety-PickUp task on a physical platform.
Experimental setup
Datasets: ProcTHOR Objaverse DivScene
Benchmarks: Safety-CHORES DivScene iTHOR ProcTHOR
Metrics: success rate cumulative cost
Baselines: SPOC FLaRe FLaRe-RS Poliformer
Builds on: SPOC FLaRe Safe-RLHF PPO constrained Markov decision process
Closest papers by meaning
- SAFE: Multitask Failure Detection for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.91)
- RedVLA: Physical Red Teaming for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
- Safe RLHF-V: Safe Reinforcement Learning from Multi-modal Human Feedback (NeurIPS 2025 · similarity 0.89)
- FOSP: Fine-tuning Offline Safe Policy through World Models (ICLR 2025 · similarity 0.89)
- Enhancing Safety in Reinforcement Learning with Human Feedback via Rectified Policy Optimization (NeurIPS 2025 · similarity 0.88)
- Failing Forward: Adaptive Failure-Informed Learning for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.