Genesis Playground is an open-source sandbox for building and sharing Robot Learning applications on top of the Genesis simulator. It provides environments, agents, and typed schemas into one cohesive workspace so researchers and students can prototype quickly and compare methods fairly.

Note: This project is for education and research. It is not an official Genesis product.

The playground is built with a modular design that separates concerns:

graph TD
    A[gs-schemas] --> B[gs-env]
    A --> C[gs-agent]

• gs-schemas - Shared data structures and interfaces
• gs-env - Robot simulation and real-world environments
• gs-agent - Robot learning algorithms
• examples - Ready-to-run examples and demos

# Install uv (fast Python package manager)
curl -LsSf https://astral.sh/uv/install.sh | sh

# Clone and setup the playground
git clone [email protected]:yun-long/GenesisPlayground.git
cd GenesisPlayground

uv sync --package gs-env

source .venv/bin/activate

python3 examples/run_ppo_gym.py # OpenAI gym inverted pendulum

or

python3 examples/run_ppo_gs.py # Genesis example

The playground includes a comprehensive teleop system for recording robot demonstrations and replaying them for behavior cloning and analysis.

Use the teleop interface to record robot demonstrations:

# Record demonstrations with Franka robot (pick cube task)
python examples/run_pick_cube_teleop.py

Controls:

• WASD: Move end-effector in X-Y plane
• QE: Move end-effector up/down (Z-axis)
• IJKL: Rotate end-effector orientation
• Space: Close/open gripper
• R: Reset scene
• Esc: Quit and save trajectory

Recording Features:

• Automatic trajectory saving to trajectories/ directory
• Timestamped commands and observations
• Optimized data storage with minimal copying
• Real-time performance monitoring

Replay recorded demonstrations for analysis or behavior cloning:

# Replay the latest recorded trajectory
python examples/replay_with_viewer.py

# Replay a specific trajectory file
python examples/replay_with_viewer.py --trajectory-file franka_pick_place_1757368298.pkl

Replay Features:

• Visual playback with Genesis viewer
• Step-by-step trajectory analysis
• Command and observation data access
• Compatible with behavior cloning algorithms
• Configurable filename prefix for different robots/tasks
• Support for replaying specific trajectory files

Train behavior cloning on gym environments:

# Train behavior cloning model on gym environment (Pendulum-v1)
python examples/run_bc_gym.py

Recorded trajectories are saved as pickle files with the following structure:

{
    "metadata": {
        "env_name": "pick_cube_default",
        "robot_type": "franka",
        "recording_duration": 45.2,
        "num_steps": 4520,
        "timestamp": "2025-01-08T15:30:45"
    },
    "trajectory": [
        {
            "timestamp": 0.0,
            "command": {
                "position": [x, y, z],
                "orientation": [qw, qx, qy, qz],
                "gripper_close": 0.0,
                "reset_scene": False,
                "quit_teleop": False
            },
            "observation": {
                "ee_pose": [x, y, z, qw, qx, qy, qz],
                "cube_pos": [x, y, z],
                "cube_quat": [qw, qx, qy, qz]
            }
        },
        # ... more steps
    ]
}

# Install pre-commit hooks
pre-commit install

# Run all checks
pre-commit run --all-files

# Run type checking
uv sync --package gs-env
pyright src/env

uv sync --package gs-agent  
pyright src/agent

• Address CI errors
• Add coding guidelines
• Add teleop interface and examples
• Add BC algorithms