🦿 DOG LAB v7 3D — “Iron Wolf” (CPG Enhanced)

A single-file experiment in 3D quadruped locomotion where a Central Pattern Generator (CPG) provides a rhythmic gait prior, and a multi-timescale “liquid reservoir” actor-critic learns residual corrections on top of that prior. Training is driven by an evolutionary outer loop that searches hyperparameters + reward weights, while the UI renders the sim in Three.js and shows live telemetry.

Primary metric: distance traveled (mean_distance)
UI: Flask server (local) + Three.js dashboard (in-browser)

Screenshot

(https://dormantone.github.io/doglab/)

Why this exists

Learning quadruped locomotion from scratch is hard and often brittle. This version makes it easier by:

• Using a CPG to supply a reasonable trotting baseline
• Having the network learn residuals, not raw joint targets
• Adding coupled torso–leg dynamics so motion looks/feels more physical
• Rewarding coordination (diagonal sync, symmetry, smoothness) and penalizing instability and lateral drift
• Running a natural-gradient-ish evolution loop with per-gene adaptive sigma + diversity injection

Features

Control + learning

• CPG Prior (trot gait, diagonal pairs synced)
• Residual policy: network outputs small adjustments on CPG targets
• Multi-timescale reservoir (“liquid” net) feeding an A2C-style actor-critic head
• Evolutionary outer loop
  • per-gene adaptive mutation (sigma) driven by sensitivity
  • BLX-α crossover
  • periodic diversity injection + stagnation boost

Physics

• 3D torso + 4 legs (12 joints) with:
  • ground contact via mirror-floor shadow anchors
  • friction-limited tangential forces
  • soft joint limits (damping near boundaries)
  • reaction torques from leg motors back onto torso

UI

• Live Three.js render (orbit camera)
• “Hyperspace” population projection view
• Neural activation view (demo mode)
• Per-gene sigma/sensitivity view
• Telemetry plots (reward, distance, height)

Requirements

• Python: 3.9+ (3.10/3.11 recommended)
• OS: macOS / Linux / Windows (works anywhere Python + Torch works)
• Compute: CPU is fine (default). Faster cores = faster generations.

Python packages

• flask
• numpy
• torch

The UI pulls Three.js from a CDN (no npm build step).

Quick start

1) Install deps

python -m venv .venv source .venv/bin/activate # macOS/Linux

.venv\Scripts\activate # Windows

pip install -U pip pip install flask numpy torch If torch install is annoying on your platform, use the official PyTorch install selector for your OS/Python and then come back.

2. Run python dog_lab_v7_3d.py

Then open: http://127.0.0.1:5005