Venue: NeurIPS 2025 (Spotlight) · Authors: Borong Zhang, Yuhao Zhang, Jiaming Ji, Yingshan Lei, Yishuai Cai, Josef Dai, Yuanpei Chen, Yaodong Yang (PKU-Alignment, Peking University) · arXiv: 2503.03480 Category: Safety / Robustness

flowchart LR
  Base[VLA training] --> CL[Constrained learning:<br/>CMDP + Integrated Safety Approach ISA]
  CL -- reward-first --> S[Reward signal: task success]
  CL -- constraints --> Sf[Safety signal: violation cost]
  S & Sf --> Balance[Solve constrained optimization]
  Balance --> VLA[SafeVLA policy]
  VLA -- eval --> R[-83.58% violations<br/>+3.85% success<br/>OOD robust]

VLAs trained by plain imitation learning occasionally take unsafe actions — collisions, excessive force, tipping, spills. There's no principled mechanism in the training objective to penalize these; they come out as rare but costly failures. LLM-alignment tools (RLHF, DPO) don't directly translate because embodied safety is about physical consequences, not text preferences.

Formulate VLA training as a Constrained Markov Decision Process (CMDP):

• Primary objective: maximize task success.
• Constraints: safety violation cost ≤ threshold.
• Solve using the Integrated Safety Approach (ISA) — a pipeline that models safety requirements, actively elicits diverse unsafe behaviors, constrains the policy via safe RL (CMDP with Lagrangian / min-max optimization against elicited risks), and assures safety through targeted evaluation.

Constraints cover an Object Safety Constraint (penalizing unintended object displacement/rotation) and a Robot Safety Constraint (preventing collisions with forbidden structures), observable during rollouts.

Built on the SPOC transformer VLA, fine-tuned and evaluated in Safety-CHORES — a new AI2-THOR / ProcTHOR-based benchmark with millions of unique scenes that extends the CHORES task suite with safety constraints.

• −83.58% cumulative safety-violation cost over SOTA baselines.
• +3.85% task success rate — safety doesn't have to cost performance.
• Generalizes to OOD perturbations (unseen objects, lighting, initial configurations).

First algorithm to explicitly incorporate safety constraints into VLAs (and first comprehensive VLA safety benchmark, Safety-CHORES) — seeds an emerging sub-literature at NeurIPS 2025 alongside SAFE (failure detection) and Latent Policy Barrier (OOD recovery).

Before NeurIPS 2025, VLA safety was scattered. After it, three complementary approaches form a proto-stack:

• Training-time: SafeVLA (CMDP)
• Detection-time: SAFE (failure classifier)
• Inference-time: Latent Policy Barrier (stay on expert manifold)

No direct 1:1 ICLR 2026 descendant yet — safety is still maturing.

• arXiv: https://arxiv.org/abs/2503.03480
• NeurIPS virtual page: https://neurips.cc/virtual/2025/poster/116975
• GitHub: https://github.com/PKU-Alignment/SafeVLA

• Latent Policy Barrier (inference-time safety sibling)
• Survey: VLA & Manipulation (ICLR 2026)

← Back to NeurIPS-2025