EMBERBench documentation
EMBERBench docs
The open task, transparent scoring and public development workflow.
Start here
4 guidesThe open task
5 guides- Simulation contractMovement, collisions, exits and fire semantics.
- Scenario schemaPublic scenario fields and validation rules.
- Public scenariosThe diagnostic scenarios available for training and development.
- Reference agentsPublic hand-coded agents and planning references.
- Development generatorOpen training worlds, separate from the official distribution.
Evaluation & scoring
7 guides- EMBER scoringThe full formula, weights, gates, aggregation and failure handling.
- Episode metricsRaw outcomes and authenticated measurements.
- Crowd diagnosticsObserved congestion and throughput signals; no crowd score.
- Running evaluationsPublic paired schedules, records and result handling.
- Frozen environmentsPublic pack identity, digests and matched fire worlds.
- Statistical analysisIntervals, resampling and paired comparisons.
- Fairness contractModel-visible information and simulator fairness.