Research

ManiGuard: Finishing the Task Is Not the Same as Doing It Safely

Robb Harlan 5 min read

A policy can put the mug on the tray and still knock the glass over on the way. ManiGuard, from Northwestern, Stanford, and William & Mary, is a benchmark that scores those as different outcomes. Safety is an LTLf formula checked by an automaton over physics predicates — not an LLM judge.

The suite has 200 locked household tasks in six families, each run in-distribution and under four single-axis shifts, for 1,000 scenarios. Across more than 23,000 evaluation rollouts, 6–21% of successful trials still violate the spec.

ManiGuard-Bench six task families with an LTL runtime monitor and safety-annotated dataset
Six families, a compiled LTLf monitor, and 8,000 safe-success demos. Source: Peng et al., arXiv:2608.17386.

What the benchmark actually checks

Families sit on a skill × constraint grid: Clutter, Lid, Cabinet, Jar, Stack, Dusty. Spatial invariants (do not topple, do not spill) sit next to ordering rules (lid on before you carry; jar closed before you lift). OmniGibson / Isaac Sim supplies contact and fluid. The same specs are scored on a physical Franka.

A rollout is “engaged” at first whole-arm contact with a task object. Idle policies look safe because they never touch anything. ManiGuard reports that separately as vacuous-safe.

The authors release 8,000 safety-annotated demonstrations — 40 per base task — filtered by the same monitor. Generation is cuRobo planning from 1,547 grasps on 221 objects, plus GELLO / SO-101 teleop. Code is on GitHub; the dataset is on Hugging Face.

Grid of ManiGuard tasks under appearance, background, location, and language shifts
One-axis OOD shifts; the safety formula stays fixed. Source: Peng et al., arXiv:2608.17386.

Fine-tuning helps. It does not finish the job.

Zero-shot π0.5, π0, and SmolVLA sit near 0% safe-success. They look 78–83% “safe” mostly by not engaging; among engaged rollouts they violate 30–54% of the time.

Supervised fine-tuning on the suite raises safe task completion to 7.5–29.8% and engaged-and-safe behavior to 51–72%. Fine-tuned π0.5 goes from 41% → 84% engagement while engaged violations fall 54% → 21%. Two of six families stay below 2% safe-success for every policy. 21–42% of engaged rollouts still violate. More of the same demos does not close that gap. The pattern holds under the OOD shifts and on the Franka.

Collage of dozens of simulated household manipulation scenes used in ManiGuard-Bench
A slice of the 200 base tasks. Source: ManiGuard project page.

A Human’s Take

Success rate as a single number has been lying to this field. If two policies tie on task success and differ by six points on violations, the leaderboard was scoring the wrong thing.

I want this monitor on hardware that can actually spill. The remaining 21–42% engaged-violation band is the real product requirement: not “can it do the task,” but “does it knock the jar over while doing it.”

Sources