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

The public main branch is the Bittle-only release (v0.8.0). The earlier Go2 (Paper 1) and Freenove (Paper 2) platforms are preserved as git tags — run git checkout v0.4.1-paper1 or git checkout v0.4.3-paper2 to reproduce those results with their original assets.

First Training Run (Bittle, simulation)

# Train the Bittle on flat ground: OpenCat Trot gait + SNN refinement.
# --neural-cpg is REQUIRED for the Bittle — it loads the OpenCat Trot controller.
# Without it the SpinalCPG path is used and the robot falls immediately.
python scripts/train_baby.py --creature-name bittle --neural-cpg \
    --scene flat --steps 25000 --hardware-sensors --fresh --snn-substeps 10

The creature walks from step 1 — the CPG provides the gait — and the SNN refines it through R-STDP as the competence gate hands control from the CPG toward the learned actor.

Analyzing Results

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

# Render a training video with dashboard overlay
python scripts/render_bittle.py creatures/bittle/<run>/training_log.bin --speed 3

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

Running on Real Hardware

The Bittle runs over a WiFi bridge from a host PC, using the same src/brain/ code as the simulator. See Hardware for the full WiFi quickstart (channel check → live gait with the SNN → live dashboard). The Freenove Robot Dog deployment on a Raspberry Pi 4 is documented there as the Paper 2 platform.

Key CLI Parameters

Parameter	Description
`--creature-name`	Robot body: `bittle` (Go2 / Freenove live on the paper tags)
`--neural-cpg`	Load the OpenCat Trot controller — required for the Bittle
`--scene`	Training scene, e.g. `flat`
`--steps`	Training steps (~25k typical for the Bittle)
`--hardware-sensors`	Use the hardware sensor layout
`--snn-substeps`	SNN integration substeps per control step
`--fresh`	Ignore any saved brain, start from scratch