Source note
OopsieVerse: A Safety Benchmark with Damage-Aware Simulation for Robot Manipulation
Summary
OopsieVerse adds explicit damage tracking to household robot simulation so policies can be judged on both task completion and physical harm. It pairs DamageSim with OopsieBench, a 32-task benchmark across OmniGibson and RoboCasa.
Problem
- Household manipulation benchmarks often score goal completion without measuring broken objects, robot damage, heat damage, or liquid damage.
- Real-world training and evaluation can be expensive and unsafe when failures damage the robot or the home.
- Existing safety costs usually need task-specific constraints, which makes cross-task comparison hard.
Approach
- DamageSim augments a POMDP with per-object and per-link health, initialized on a 0 to 100 scale and reduced when damage evaluators fire.
- Mechanical damage uses simulator contact forces and link acceleration to estimate impulsive and sustained load, then subtracts health when load passes an object-specific threshold.
- Thermal damage subtracts health when object temperature exceeds hot or cold thresholds; the excerpt says this is implemented only in OmniGibson because it needs temperature state.
- Fluid damage subtracts health when liquid contact passes a threshold; the excerpt says this is implemented only in OmniGibson because it needs fluid particles.
- OopsieBench provides household tasks with unsafe shortcuts and safer alternatives, so users can measure success and damage separately.
Results
- OopsieBench contains 32 task instances, 21 unique task designs, 17 OmniGibson tasks, and 15 RoboCasa tasks.
- DamageSim is implemented in 2 physics stacks: OmniGibson on Nvidia Omniverse and RoboCasa/Robosuite on MuJoCo.
- The damage model covers 3 main damage classes: mechanical, thermal, and fluid; mechanical damage includes impact, compression, and tension examples.
- The excerpt claims live damage feedback makes human teleoperation demonstrations safer, but it gives no exact metric or percentage.
- The excerpt claims damage-conditioned imitation learning, damage-aware reinforcement learning, VLA policy benchmarking, and sim-to-real safety gains, but it provides no quantitative policy results in the supplied text.