ManiGuard benchmark reveals safety gaps in robotic manipulation policies
Even successful robot tasks fail safety checks 6-21% of the time.
ManiGuard is a new specification-grounded framework from researchers at Northwestern, UCLA, and UCSD for evaluating and improving the safety of robotic manipulation policies powered by foundation models. The benchmark comprises the ManiGuard-Bench task suite, which organizes six contact-rich household task families (e.g., tool use, cleaning, cooking) into 200 locked base tasks along a skill × constraint taxonomy. Each task is evaluated under one in-distribution and four single-axis out-of-distribution perturbations, yielding 1,000 locked scenarios. Safety is specified independently of task success, and every rollout is runtime-checked by LTLf-grounded automaton monitors over physics-based predicates—not learned classifiers or LLM judges—both in simulation and on a physical Franka robot platform.
The authors also built a paired safety-annotated trajectory-generation pipeline that combines automated motion-planning generation with human teleoperation. This pipeline supports safety-aware fine-tuning and releases 8,000 demonstrations (40 per base task). Benchmarking zero-shot and fine-tuned vision-language-action models (VLAs) across more than 23,000 rollouts revealed three key findings: (1) safety must be measured independently of task success, since 6-21% of successful rollouts violate the specification; (2) fine-tuning on the ManiGuard suite improves safe task completion from near zero to 7.5-29.8% and boosts engaged-and-safe behavior from 16-40% to 51-72%; (3) a persistent gap remains—scaling demonstrations doesn't close it, with 21-42% of engaged rollouts still violating safety, and two of the six task families achieving under 2% safe success for every policy. These failures persist under distribution shift and on physical hardware.
- ManiGuard includes 200 locked base tasks, 1,000 scenarios, and 8,000 safety-annotated demonstrations using LTLf-based runtime monitors.
- Across 23,000+ rollouts, 6-21% of successful VLA rollouts violate safety specifications, showing task success ≠ safety.
- Fine-tuning on ManiGuard raises safe completion from ~0% to 29.8% but leaves 21-42% of engaged rollouts unsafe, with two task families below 2% safe success.
- Evaluations run on both simulation and physical Franka platform, including out-of-distribution perturbations.
Why It Matters
ManiGuard provides a rigorous, specification-based safety benchmark that reveals critical weaknesses in robotic manipulation policies, guiding safer real-world deployment.