g1_crawl

g1_crawl copied to clipboard

G1 Crawl (Isaac Lab)

Note: I put the final policies in sim2sim_mj/policies but I dont recall the exact reward terms that resulted in them (especially the final crawl), so you might have to do some sweeps to recreate the results seen in video.

Prerequisites & Setup

• Isaac Lab environment: Follow the Isaac Lab documentation to install and create a conda environment. Use that conda env for this repo.
• This project assumes a conda env named g1-crawl. Always run commands like: conda activate g1-crawl && python ....
• Optional (for sweep video post-processing): ensure ffmpeg is installed and on PATH.

Train a Policy

• Minimal headless training:

conda activate g1-crawl && \
python scripts/rsl_rl/train.py --task g1-crawl --headless

• Common overrides (examples):

conda activate g1-crawl && \
python scripts/rsl_rl/train.py --task g1-crawl --headless \
  agent.max_iterations=2500

• Logs and checkpoints are written under logs/rsl_rl/<experiment_name>/<timestamp_UUID>/.

Play a Trained Policy

• Play latest checkpoint from a run directory:

conda activate g1-crawl && \
python scripts/rsl_rl/play.py --task g1-crawl --headless --video --video_length 200 --enable_cameras \
  agent.experiment_name=<your_experiment> \
  agent.load_run=<timestamp_UUID>

• Play a specific checkpoint by absolute path:

conda activate g1-crawl && \
python scripts/rsl_rl/play.py --task g1-crawl --headless --video --video_length 200 --enable_cameras \
  --checkpoint /abs/path/to/model_XXXX.pt

Run a Parameter Sweep (run_sweep.py)

1. Open run_sweep.py and edit the first section:
   • SWEEP_QUEUE: add/adjust entries and parameters to sweep (use lists for values).
   • Use the special string "__OMIT__" to test “not setting” a parameter.
   • Set AUTO_SUSPEND=False if you do not want the machine to sleep after the sweep.
2. Run the sweep:

conda activate g1-crawl && \
python run_sweep.py

3. What it does:
   • Trains for each combination, then auto-plays and saves videos.
   • Writes training logs to logs/rsl_rl/<experiment_name>/....
   • Writes sweep logs and concatenated videos to sweep-logs/<experiment_name>_<timestamp>/.
   • If ffmpeg is available, also creates a labeled concatenated video and a segment mapping.

Notes:

• resume_checkpoint (when included in a sweep combo) is passed as a CLI flag automatically.
• The script fails loudly on missing run directories or file I/O errors so issues are visible immediately.

Sim2Sim Standalone Playback (sim2sim_mj/run_sim2sim.py)

This runs exported policies in standard MuJoCo (no Isaac Lab runtime needed during playback).

1. Dependencies (within the same conda env):

conda activate g1-crawl && \
pip install -r sim2sim_mj/requirements.txt

2. Assets:
   • Ensure sim2sim_mj/scene.xml exists (export your model if needed).
   • Place policy files under sim2sim_mj/policies/ and update the CONFIG in sim2sim_mj/run_sim2sim.py:
     • policy_cycle: list the .pt files you want to cycle through.
     • policy_defaults: map each policy path to "stand" or "crawl" for the default joint set.
3. Run:

conda activate g1-crawl && \
python sim2sim_mj/run_sim2sim.py

4. Controls:
   • A single face button cycles policies (see CONFIG.cycle_button).
   • Start/Menu exits (see CONFIG.exit_button).
   • Analog sticks provide velocity commands if a standard-mapped gamepad is connected.
   • Optional CSV diagnostics are written under outputs/diagnostics/ when enabled.

Real Robot Deployment (deployment/)

• The deployment/ folder contains scripts/configs intended for running policies on the physical robot.
• These expect robot-side connectivity, drivers, and safety layers (e.g., actuator PD settings, watchdogs).
• Adjust gains, safety thresholds, and device interfaces there to match your hardware before use.
• Use with caution; validate in simulation first.

Common Paths

• Training logs/checkpoints: logs/rsl_rl/<experiment_name>/<timestamp_UUID>/
• Sweep outputs (guides/videos/mapping): sweep-logs/<experiment_name>_<timestamp>/
• Sim2Sim policies: sim2sim_mj/policies/
• Sim2Sim diagnostics: outputs/diagnostics/

3D Printed Parts (Onshape links)

• Head parts (requires heat-set inserts for screws)
• Custom arms (requires a squash ball)
• Corresponding 3MF print files live in 3d-printed-parts/.