Documentation index¶
Start with setup, then dataset → training → simulation; use hardware and bring-up when moving to the arm.
| Document | Scope |
|---|---|
| setup.md | Docker and host Python setup; contract config overview |
| dataset.md | Test rosbag2 generation, synthetic data, MuJoCo expert collection, LeRobotDataset conversion and QA |
| training.md | ACT training, checkpoints, and inference smoke checks |
| real_dataset_so101_pick_sponge_main_20261023.md | Real pick-sponge main (clean28) dataset summary and overnight ACT notes (2026-10-23, facts) |
| ros2_inference.md | ROS 2 Humble nodes, mock inference, Rerun visualization |
| simulation.md | MuJoCo closed-loop eval, interactive viewer, video export |
| evaluation.md | Planar position / yaw / success aggregation (eval_alignment) API |
| real_eval_protocol.ja.md | Real ACT open-loop scoring and live rollout evaluation protocol (Japanese) |
| units.md | Sim/contract radians vs real LeRobot degrees and gripper 0–100 |
| sim_to_real_act.md | Phase 2-2: sim-trained ACT on real SO-101 (dry-run / slow live) |
| results/README.md | Closed-loop eval logs and pending placeholders (including post50) |
| figures/README.md | Regenerating report figures (loss curve, histograms, initial XY, rollout) |
| e2e.md | End-to-end pipeline from data collection through inference |
| hardware.md | SO-ARM101 / Feetech udev, connection dry-run, current table layout (2026-10-26), purchased gear |
| real_robot_bringup.md | Mac / LeRobot 0.6.1 SO-101 Leader/Follower bring-up, current table layout (2026-10-26), session / live-eval checklist, pre-record preflight (scripts/preflight_real_so101.sh), demo recording (scripts/record_real_so101.sh), post-record QC (scripts/qc_teleop_dataset.py) |
| real_robot_deploy.md | Real-trained ACT / SmolVLA dry-run → live (no unit conversion; safety clamps) |
| ci.md | GitHub Actions CI and local lint/test |
| submission.md | OmniCampus deliverables (PDF report, GitHub, YouTube) and checklist |