What changed
Robot-manipulation papers usually reward a policy for reaching an end state. ManiGuard, a preprint published August 18, asks a second question: did the robot reach the goal without violating a stated safety condition? The researchers' benchmark contains 200 locked, contact-rich household tasks, each tested across four controlled observation shifts. A finite-trace linear temporal-logic monitor grades violations from simulated state rather than a learned or language-model judge.[1]
The metric that changes the picture
Across more than 23,000 author-run evaluation rollouts, the team reports that 6 to 21 percent of successful fine-tuned vision-language-action policy runs still broke the specified rule. That is the decision delta: task completion can mask a spill, collision, bad ordering or another encoded violation. The authors also find that raw safety alone can reward inaction, because zero-shot policies' apparent safety often came from runs that never engaged the task object.[1]
What is real—and what is not
The study includes a physical Franka comparison, but it is an initial transfer test, not a deployment safety case. It covers 12 matched simulation-and-hardware cells with 30 rollouts each: 360 hardware attempts by calculation. The paper reports that simulation ranked policies similarly on task success and safe success, but did not predict the violation rate; its limitations state that physical verdicts were operator-scored, not generated by the state-based monitor used in simulation.[1]
The reproduction test
The release has practical artifacts: the project docs give an evaluation recipe, while a public CC BY 4.0 Clutter package contains 2,200 five-camera episodes across 55 tasks. But the code URL linked from both paper and docs returned HTTP 404 on August 19, so independent end-to-end reproduction cannot yet be assumed. The next useful checkpoint is a restored repository and third-party tests on another arm, sensor set and workcell that publish unsafe-completion rates beside task success.[1,2,3,4]