CPG + RL Swimming Robot

Hybrid control for a swimming robot: CPG (Central Pattern Generator) and RL (PPO). The simulator is built on Genesis SPH and is configured to run on CUDA and macOS (Metal) as well as other platforms.

Features

• Hybrid Control: CPG + PPO modulation
• Paddle Disturbance (enabled): deterministic, controllable wave generator
• Genesis SPH Simulation: high-fidelity water simulation (backend auto-selected)

Installation

Prerequisites

• Python 3.11 recommended (Conda works well on macOS)
• Genesis: genesis-world

Install dependencies

pip install -r requirements.txt

If you use Conda on macOS and want to be explicit:

/opt/anaconda3/bin/python3 -m pip install -r requirements.txt

Quick Start

Training

No paddle disturbance:

/opt/anaconda3/bin/python3 run.py --mode cpg_rl --cpg 0 --profile short

Enable paddle disturbance:

/opt/anaconda3/bin/python3 run.py --mode cpg_rl --cpg 0 --profile short \
  --paddle --paddle-count 2 --paddle-freq 0.2 --paddle-amp 0.17

Key flags

• --mode: direct | cpg | cpg_rl | res_rl
• --cpg: 0 = PhaseOscillator (recommended), 1 = Matsuoka
• --paddle: enable paddle disturbance
• --paddle-count: number of paddles
• --paddle-freq: oscillation frequency (Hz)
• --paddle-amp: oscillation amplitude (m)

Platform notes

Genesis backend selection

The simulator auto-selects a safe backend (e.g., Metal on macOS). You can override it:

GENESIS_BACKEND=metal /opt/anaconda3/bin/python3 run.py --mode cpg_rl --profile short

Supported values: cpu, metal, cuda (if available).

Project structure

.
├── run.py
├── eval.py
├── cpg_ppo_agent.py
├── cpg_env_adapter.py
├── cpg_config_utils.py
├── env/
│   ├── cpg.py
│   └── swim_gs_env.py
└── eelrobotv2_urdf/
    ├── urdf/
    └── meshes/

Troubleshooting

First run is slow

Genesis may compile kernels on the first run (especially on macOS/Metal). Let it finish; subsequent runs are faster.

wandb not installed

This repo treats wandb as optional. If you want W&B logging:

pip install wandb