PhysSafe is an open benchmark for evaluating physical safety in humanoid robots.
The benchmark provides standardized safety-critical scenarios, evaluation metrics, and trajectory data formats for reproducible research in robot safety.
Humanoid robots are increasingly deployed in environments where safety is critical, including:
- human–robot interaction
- collaborative manipulation
- dynamic locomotion
- service robotics
However, existing benchmarks focus primarily on task performance, while physical safety remains under-evaluated.
PhysSafe aims to provide the first benchmark specifically designed for humanoid physical safety evaluation.
PhysSafe currently includes the following safety scenarios:
Robot operates near humans while maintaining safe distance.
Safety risks:
- collision
- unsafe proximity
- unexpected contact
Humanoid walking on irregular terrain.
Safety risks:
- fall
- slip
- instability
Robot experiences unexpected pushes.
Safety risks:
- loss of balance
- unsafe recovery
- collision
PhysSafe evaluates controllers using safety metrics:
| Metric | Description |
|---|---|
| Collision Rate | fraction of episodes with collision |
| Fall Rate | fraction of episodes with falls |
| Safety Violation Rate | fraction of episodes violating safety constraints |
| Task Success Rate | fraction of successful task completion |
| Intervention Rate | frequency of safety controller intervention |
| Average Safety Margin | average distance to safety boundary |
Example evaluation:
python evaluation/evaluate.py \
--hdf5 sample_rollouts.h5 \
--scenario_id HRI_HUMAN_PROX_001 \
--level L2 \
--out result.json