Repository Map¶

AM-Bench separates simulation, imitation learning, command-line workflows, internal engineering notes, and third-party integrations.

am_isaac/
  source/am_isaac/          simulation extension
  source/am_isaac_il/       learned-policy and data adapters
  scripts/environments/     registration, smoke tests, teleoperation
  scripts/data/             recording, validation, format conversion
  docs/                     public documentation
  ext/                      optional external dependencies

Simulation package¶

Path	Responsibility
`tasks/`	Scene setup, observations, rewards, success, and reset logic
`tasks/base/`	Shared `BaseEnv`, configuration, registrations, and robot profiles
`scenes/`	Reusable task-agnostic scene components
`robots/`	Morphology, assets, rotor layouts, and `RobotIO` semantics
`controllers/`	Action pipelines, IK integration, wrench controllers, and allocation
`disturbance/`	Aerodynamic and disturbance models
`policies/scripted/`	Task-aware scripted policies
`recording/`	Canonical LeRobot and video writers
`evaluation/`	Controller tracking and limit-use diagnostics

Imitation-learning package¶

Path	Responsibility
`data/`	Logical resampling and action-semantic validation
`pipeline/`	Dataset and action transformations shared by learned policies
`eval/`	Common evaluator CLI, batching, outputs, and tracking lifecycle
`policies/act/`	ACT datasets, processor, model, training, and evaluation
`policies/dp/`	Diffusion Policy integration and UMI dependency
`policies/pi/`	OpenPI client-side evaluation adapter

The ext/pyroki and ext/acados trees are external integrations. Changes to them should be isolated from ordinary task or documentation work.

Development-only datasets, checkpoints, videos, experiment wrappers, and generated build outputs are not part of the public package boundary.