Complete Workflow
MH-FLOCKE provides a full pipeline from training a quadruped creature to producing publication-quality videos with data-driven sonification. Every step operates on FLOG binary logs.
Train ──→ FLOG ──→ Render (Video)
├──→ Sonify (Audio)
├──→ Dashboard (Analysis)
└──→ Behavioral Breakdown
Step 1 — Train
Train a Unitree Go2 quadruped. The creature starts with an innate CPG gait and learns through R-STDP.
python scripts/train_v032.py \
--creature-name go2 \
--scene "walk on flat meadow" \
--steps 50000 \
--skip-morph-check \
--no-terrain \
--auto-reset 500 \
--seed 42
Key arguments:
--creature-name go2— Unitree Go2 (12 actuators)--steps— Training steps (50k ≈ 90 min on CPU)--no-terrain— Flat ground--auto-reset 500— Reset if fallen--seed— Random seed
Output: creatures/go2/v034_TIMESTAMP/ with training_log.bin (FLOG), snn_state.pt, checkpoint.pt, knowledge.json.
Step 2 — Render Video
Render 1440p video from FLOG with dashboard overlay.
python scripts/render_go2_mujoco.py \
--flog creatures/go2/v034_TIMESTAMP/training_log.bin \
--overlay --speed 2
Dashboard overlay: Brain3D, CPG/Actor balance, Cerebellar PE, Behavioral state, Neuromodulation, Distance/Velocity.
Step 3 — Sonify
Data-driven audio from FLOG. Every sound = real training metric.
python scripts/sonify_flog.py \
--flog creatures/go2/v034_TIMESTAMP/training_log.bin \
--speed 2
Audio layers: SNN Crackle, Motor Hum, Heartbeat Pulse, Cerebellum Tone, Scent Chime, Fall Impact, DA Melody, Ambient Pad, Ball Proximity.
Mux with video: --mux output_video.mp4
Step 4 — Dashboard
python flog_server.py # Open http://localhost:5050
6 charts: distance, velocity, falls, CPG/actor weight, cerebellar correction, behavioral timeline.
Step 5 — Behavioral Analysis
python scripts/behavioral_breakdown.py
Behavior time per condition, aggregated across seeds.
Ablation
# A1 — CPG only
python scripts/train_v032.py --creature-name go2 \
--scene "walk on flat meadow" --steps 50000 \
--no-terrain --auto-reset 500 --seed 42 --ablation cpg
# B1 — SNN+Cerebellum (default)
python scripts/train_v032.py --creature-name go2 \
--scene "walk on flat meadow" --steps 50000 \
--no-terrain --auto-reset 500 --seed 42
# PPO baseline
python scripts/ppo_baseline_go2.py --steps 50000 --seed 42
Seeds: 7, 42, 55, 123, 256, 314, 808, 999, 1337, 2025.
System Requirements
- Python 3.11+
- MuJoCo ≥ 3.0
- PyTorch (CPU sufficient)
- Pillow, msgpack, FFmpeg
- RAM: 8 GB min
- Training: ~90 min / 50k steps