Choose a Workflow¶
The first simulation proves that the runtime can create and step an AM-Bench task. Choose the next page by the outcome you want.
| Goal | Continue with |
|---|---|
| Understand benchmark scope | Benchmark Design |
| Choose a task | Task Suite |
| Compare robot embodiments | Robot Platforms |
| Collect scripted or teleoperated data | Collect Demonstrations |
| Validate an existing dataset | Validate Datasets |
| Train ACT, DP, PI0, or PI0.5 | Policy Guides |
| Interpret evaluation artifacts | Evaluation Outputs |
| Add a task, robot, controller, or policy | Extend AM-Bench |
Recommended progression¶
For a new task or platform, use the smallest diagnostic ladder:
- run the EE-only PID variant;
- verify the scripted policy and success criteria;
- collect and validate a small dataset;
- train or load one policy baseline;
- evaluate through the standardized artifact path;
- change one platform, controller, interface, or physical effect at a time.
This progression keeps task-logic failures separate from policy, controller, and actuation failures.