How sgyliu8/lerobot works
PhysicalAI SO101 Lab is a local, upstream-first robot learning workbench. Its principal workflow is supervised Leader-to-Follower demonstration with two cameras, followed by recording review and audit, optional data preparation, local ACT training, checkpoint recovery, and offline output checks before any supervised physical trial. The README documents these capabilities, but most workbench, device, recording, and training implementation belongs to the LeLab/LeRobot dependency and is not present in the supplied tree; those links reflect documented behavior. Sampled project code supports the color-task evidence/split path and ROS 2 read-only replay.
Main entry points: training_split.py Split binding · color_sorting_cli.py Evidence CLI · joint_state_replay.py ROS state replay · color_sorting.py Labels and splits
Components 19
Capture and review
- LeLab workbench
- Leader arm
- Follower arm
- Dual cameras
- Episode recording
- Browse and review
- Dataset audit
tools/audit_dataset.py
Learning workflow
- ACT training
- Split binding
so101_lab/training_split.py - Checkpoints
- Offline output check
Color task evidence
- Color task profile
so101_lab/color_sorting.py - Evidence CLI
so101_lab/color_sorting_cli.py - Labels and splits
so101_lab/color_sorting.py
Optional integrations
- ROS state replay
integrations/ros2/joint_state_replay.py - ROS description prep
integrations/ros2/prepare_description.py
Other components
- Operator
- ROS 2
- Local dataset
Connections 26
- Operator → LeLab workbench: operates
- Operator → Leader arm: demonstrates
- LeLab workbench → Leader arm: reads motion
- LeLab workbench → Follower arm: dispatches control
- LeLab workbench → Dual cameras: captures video
- LeLab workbench → Episode recording: saves episodes
- Episode recording → Local dataset: persists data
- Operator → Browse and review: reviews
- Browse and review → Local dataset: reads episodes
- Dataset audit → Local dataset: audits read-only
- Operator → ACT training: configures
- ACT training → Split binding: binds color split
- Split binding → Local dataset: scopes episodes
- ACT training → Local dataset: reads training data
- ACT training → Checkpoints: saves state
- Checkpoints → ACT training: resumes in new job
- ACT training → Offline output check: produces policy
- Operator → Evidence CLI: runs commands
- Evidence CLI → Color task profile: validates profile
- Evidence CLI → Labels and splits: writes labels and splits
- Color task profile → Labels and splits: defines task contract
- Split binding → Labels and splits: reads frozen manifest
- Operator → ROS state replay: starts replay
- ROS state replay → Local dataset: reads trajectory
- ROS state replay → ROS 2: publishes state
- ROS description prep → ROS 2: prepares description assets