Embodied planners can compose subtasks whose interactions create physical risks that isolated safety checks miss. PlanGuard is a pre-execution detector that evaluates the whole multi-step plan against the current visual environment. It was trained and evaluated on MSP-Safe, a new dataset built from paired task construction, diverse plan generation, and three-judge safety annotation. The authors first use task-oriented supervised fine-tuning, then apply STAC-OPD, a distillation method where a strong teacher provides adaptive supervision along student on-policy trajectories. The resulting PlanGuard-2B reaches average 87.15% accuracy and 87.21% F1 across all test subsets. The interest is that compact models can catch compositional physical risks before execution, closing a gap left by general guardrails and single-subtask detectors.
PlanGuard detects physical risks across complete embodied-agent plans
A compact detector evaluates planned action sequences against the visual environment before an embodied agent begins execution.
Chinese Tech
Junchi Chen · Changtao Miao · Yuxiao Xiang · Zhenchao Jin · Haojie Yuan · Qi Chu · +6 more
Anhui Province Key Laboratory of Digital Security · The University of Hong Kong · Ant Digital Technologies · Ant Group
Research Digest··2 min read
Thread:Embodied Plan Safety
In one line
PlanGuard detects physical safety risks in complete multi-step embodied plans before execution, reaching 87.15% accuracy and 87.21% F1 at 2B scale.
Before you read this
Three ideas this paper builds onagent safetyRisks from AI agents including timing errors, shutdown avoidance, collusion, overclaiming, and unsafe execution, mitigated by constrained decoding, safety harnesses, and swarm governance.188 papers →on-policy distillationA training method where a student model learns from a teacher using its own current policy's outputs, often improving reasoning and instruction following.131 papers →synthetic data generationCreating training examples artificially, often using an LLM, to overcome real-data scarcity or privacy constraints.91 papers →
Read it as
Chen and colleagues introduce PlanGuard, a guardrail designed to identify physical hazards arising from dependencies among multiple planned actions, rather than assessing each action separately.
The authors constructed MSP-Safe using paired tasks, plans produced by multiple planners, and safety labels from three judges.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.