Dataset Format¶

The canonical AM-Bench data boundary is a finalized LeRobot 0.4.4 dataset produced by the repository recorder. Policy-specific zarr or OpenPI layouts are derived artifacts, not source datasets.

Session layout¶

<session-root>/
  env_cfg.yaml
  lerobot/
    meta/info.json
    meta/episodes/
    data/
    images/ or videos/
  videos/                 # optional recorder videos

env_cfg.yaml records the task configuration used for collection. The lerobot/ directory is the canonical dataset root accepted by the validator and training adapters.

Required frame features¶

Feature	Type	Meaning
`observation.state`	float32 vector	Ordered concatenation of the recorder's `--state_keys`
`action`	float32 vector	Environment action at the recorded step
`task`	string	Language instruction stored for the episode/frame
`observation.images.<camera>`	RGB image	One feature for each recorded camera

The default state keys are ee_pos ee_quat gripper_width. BaseJoint datasets normally use base_pos base_quat arm_joint_pos gripper_width.

AM-Bench metadata¶

meta/info.json includes the standard LeRobot counts, FPS, robot type, and feature declarations. Its am_isaac block records:

action_semantics: ee_absolute, ee_delta, or base_joint_absolute;
state_keys: the ordered keys concatenated into observation.state.

Canonical relative-policy training uses absolute-action sources. EE absolute actions contain 8 values; the current four-joint BaseJoint platforms contain 12. Quaternions are WXYZ.

Finalization and validation¶

Successful episodes receive finalized Parquet metadata under meta/episodes/. An interrupted collection that never finalized is not a valid training source. Run Validate Datasets before conversion or training.

Keep the canonical source immutable. DP conversion writes UMI zarr, and OpenPI export writes a pinned reader-compatible LeRobot layout plus norm-stat inputs. Regenerate either derived format from the canonical dataset when adapters change.