Getting Started

Requirements

Python 3.11+
MuJoCo (via pip install mujoco)
PyTorch
NumPy, msgpack

Installation

git clone https://github.com/MarcHesse/mhflocke.git
cd mhflocke
pip install -r requirements.txt

First Training Run

Go2 (Unitree, simulation only)

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

Expected: Go2 walks ~25m in 50k steps, 0 falls. Takes ~20 min on CPU.

Freenove (simulation with hardware drift)

python scripts/train_baby.py \
    --creature-name freenove \
    --phototaxis \
    --steps 50000 \
    --no-llm \
    --hardware-sensors \
    --no-vision \
    --auto-reset 500 \
    --fresh \
    --drift-profile creatures/freenove/drift_profiles/measured_marc_01.json

Expected: Freenove walks ~6.5m, finds 3 light targets, 0 falls. Takes ~18 min on CPU.

Analyzing Results

# FLOG dashboard (web-based)
python flog_server.py
# Open http://localhost:5050

# Render training video
python scripts/render_freenove.py creatures/freenove/v043_.../training_log.bin

# Add data-driven audio (sonification)
python scripts/sonify_flog.py --flog creatures/.../training_log.bin --speed 2 --mux output.mp4

Running on Real Hardware

See Hardware for Freenove Robot Dog deployment on Raspberry Pi 4.

Key CLI Parameters

Parameter	Script	Description
`--creature-name`	Both	Robot: `freenove` or `go2`
`--steps`	Both	Training steps (50k typical)
`--fresh`	Both	Ignore saved brain, start from scratch
`--phototaxis`	train_baby.py	Enable light-tracking navigation
`--drift-profile`	train_baby.py	Inject hardware drift into simulation
`--hardware-sensors`	train_baby.py	Use Bridge v2.5 sensor layout (48ch)
`--auto-reset`	Both	Reset after N stuck steps
`--seed`	train_v032.py	Random seed for reproducibility