Simulated robots sometimes completed tasks while violating safety requirements
Source check: Source checked. Not independent scientific confirmation.
Finding: In controlled simulations across six manipulation tasks, 82 valid executions completed their tasks while violating at least one separately checked safety rule. The tests evaluated learned ACT and RDT control policies under deliberately introduced faults. Counts refer to simulated executions in a focused test campaign.
Technology: ACT; RDT robotic manipulation policies; EAST digital-twin testing
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
- AI contribution
- Two learned manipulation policy families, ACT and RDT, produced execution traces in which task completion and predefined safety rules were evaluated separately.
- Test and result
- Five campaigns used the simulator to verify that test conditions occurred and evaluated task completion separately from predefined safety requirements.
- Limitation / still unknown
- The cases may share tasks and scenarios. This fault-focused simulation does not estimate physical incident rates. Earlier safety benchmarks already reported successful-but-unsafe robot trajectories; this result is a further specific policy/task validation, not discovery of that general gap.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.34358v1 · https://arxiv.org/html/2610.34358v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 21 Nov 2026. Review scope: v1 design, independent runtime outcome checks, equal-budget campaigns and unsafe-success results.
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 1/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.