Command overview

Every pipeline stage and tool runs as a Python module: python -m so101_mvbench.<module>. There are no console-script shims, so what you see below is the complete surface. Every command accepts --help without booting the simulator, with one exception: list_envs has no argument parsing at all and starts Isaac Sim on import, because listing the registered tasks is what it does. Give it no flags.

Commands marked sim need Isaac Sim and a GPU. The others run anywhere the package is installed.

How to read the commands

Every command in this documentation is written to be pasted into a terminal and run as it stands, from the repository root.

Where a command needs a value from you, the block sets it as a shell variable on its first lines, marked with # adjust:. Change those lines, then paste the rest unchanged:

# adjust: the bin you recorded
BIN=datasets/01_raw_gamepad/bin_c1_r1

python -m so101_mvbench.recording.trajectory_extractor \
    --dataset_root "$BIN" --output_dir "$BIN/trajectories"

Nothing else in a command needs editing. Anything adjustable is a variable at the top of its block, so a block never contains a placeholder that would break the shell if you pasted it unchanged.

Paths that appear literally, such as datasets/01_raw_gamepad/bin_c1_r1, are the conventional locations this documentation uses throughout. They work as they stand if you followed the previous stage.

Pipeline

Command	Stage	Sim	Purpose
python -m so101_mvbench.recording.teleop_recorder	1	yes	Record demonstrations (gamepad or scripted sine) into a LeRobot dataset with per-episode scene state and bin map.
python -m so101_mvbench.recording.trajectory_extractor	2a	no	Extract per-episode joint trajectories to .npy, decoupling replay from the LeRobot import chain.
python -m so101_mvbench.recording.multicam_replay	2b	yes	Re-render recorded episodes from every camera of a scene config, one pass, one dataset per bin.
python -m so101_mvbench.assembly.bin_fusion	2c	no	Fuse per-bin datasets into one trainer-ready dataset.
python -m so101_mvbench.training.train_act_camera_perm	3	no	Train an ACT policy on a chosen camera subset (wraps lerobot_train).
python -m so101_mvbench.training.checkpoint_selector	3b	no	Rank checkpoints of a finished run by evaluation loss.
python -m so101_mvbench.evaluation.async_eval	4	yes	Evaluate a policy over N parallel environments. With --full_report it also writes the files stage 5 reads.
python -m so101_mvbench.evaluation.report	5	no	Plots and a written summary for a single evaluation run.
python -m so101_mvbench.evaluation.aggregate_report	5	no	Cross-run statistical aggregation.

Audit tools

Command	Sim	Question it answers
python -m so101_mvbench.audit.determinism	yes	Does moving a camera change the physics? (If yes, every camera ablation is invalid.)
python -m so101_mvbench.audit.cube_replay	yes	Do episodes diverge across parallel environments?
python -m so101_mvbench.audit.cube_plot	no	Plot the divergence data produced by cube_replay.

Dataset tools

Command	Sim	Purpose
python -m so101_mvbench.tools.list_envs	yes	Print the registered LiftCube-* task IDs.
python -m so101_mvbench.tools.validate_bin	no	Integrity check for a recorded bin: parquet readability, MP4 presence, scene-state coverage.
python -m so101_mvbench.tools.delete_episodes	no	Remove episodes from a LeRobot dataset, renumbering safely.
python -m so101_mvbench.tools.drift_configs	no	Write scene configs with one camera coordinate shifted, for a drift sweep (What happens when the external camera moves?).

Running a whole campaign

scripts/run_campaign.sh drives stage 4 across N policies x M bins x R same-seed repeats, one Isaac boot per policy, resume-safe. Data locations come from three required environment variables:

POLICY_ROOT=outputs \
DATASETS_ROOT=datasets/01_raw_gamepad \
OUT_ROOT=outputs/campaign \
bash scripts/run_campaign.sh

DRY_RUN=1 prints the per-policy commands without booting Isaac. See the header comment of the script for every tunable.