Supported Policies¶
AM-Bench provides task-aware scripted policies and three maintained learned-policy integrations.
| Policy | Best use | Training data | Evaluation process |
|---|---|---|---|
| Scripted | task validation and demonstration collection | none | in simulator |
| ACT | compact task-specific imitation baseline | canonical LeRobot | in simulator process |
| Diffusion Policy | task-specific image-conditioned diffusion baseline | derived UMI zarr | in simulator process |
| PI0 / PI0.5 | pretrained VLA adaptation and language-conditioned evaluation | pinned OpenPI export | separate policy server and simulator client |
Shared contract¶
All learned policies begin from a validated canonical dataset and must match the environment's action semantics. Evaluators share rollout controls, progress output, artifact naming, tracking, and video behavior. Start with Evaluate a Supported Policy, then use the family guide for training and policy-only arguments.
Start small¶
Before a full experiment:
- validate the source dataset at the intended policy rate;
- run a bounded training smoke that saves a loadable checkpoint;
- evaluate one rollout with one environment and video enabled;
- inspect
results.txt,eval_summary.json, tracking analysis, and all requested camera videos; - scale only after the complete pipeline passes.
Training success does not prove evaluation compatibility, and a successful one-rollout smoke does not measure policy quality.