🐋 What is OrcaGym¶
OrcaGym is an open-source, cloud-native robot simulation platform that provides robot simulation environments that are fully compatible with the OpenAI Gym / Gymnasium interface.
In a Nutshell¶
OrcaGym = Gymnasium API + MuJoCo Physics Engine + Distributed Communication + OrcaStudio/OrcaLab Cloud Platform
Core Positioning¶
Traditional robot simulation solutions often face trade-offs between fidelity and computational efficiency. OrcaGym bridges this gap through:
- Standardized Interface: Fully compatible with the Gymnasium API, enabling zero-cost migration of existing RL algorithms
- Multiple Physics Backends: The core package locally supports MuJoCo/MuJoCoWarp/Euler; PhysX and ODE can be accessed via OrcaStudio/OrcaLab
- Cloud-Native Architecture: Enables hybrid local/remote deployment
- Realistic Rendering: Ray tracing provides high-quality observations for visual RL tasks (provided by the OrcaStudio/OrcaLab server)
Key Features¶
🎮 Gymnasium API Compatibility¶
import gymnasium as gym
env = gym.make("YourEnv-v0", frame_skip=5, orcagym_addr="localhost:50051")
obs, info = env.reset()
action = env.action_space.sample()
obs, reward, terminated, truncated, info = env.step(action)
Seamlessly integrates with mainstream RL libraries such as Stable-Baselines3, RLlib, and CleanRL.
⚡ Multiple Physics Backends¶
| Backend | Availability | Characteristics | Use Case |
|---|---|---|---|
| MuJoCo | Core package (local) | High-precision rigid-body dynamics (CPU) | Legged robots, robotic arm manipulation |
| MuJoCoWarp | Core package (local) | GPU (CUDA) accelerated parallelism | Large-scale parallel training |
| Euler | Core package (local) | Multi-vendor GPU multi-physics engine | High-fidelity multi-physics simulation |
| PhysX | Remote only (OrcaStudio/OrcaLab) | GPU accelerated, massively parallel | Swarm simulation, complex scenes |
| ODE | Remote only (OrcaStudio/OrcaLab) | Open-source, general-purpose | Rapid prototyping, educational use |
🌐 Distributed Deployment¶
Supports flexible deployment ranging from local development to large-scale remote training: - Local Mode: Run MuJoCo directly within the Python process, suitable for development and debugging - Remote Mode: Connect to a remote server for physics computation and rendering, suitable for large-scale training
📷 Sensors and Perception¶
- IMU (accelerometer, gyroscope)
- Force/torque sensors
- RGB-D cameras
- Contact force sensors
🤖 Multi-Agent Support¶
Natively supports homogeneous/heterogeneous multi-agent scenarios, where agents can operate independently or collaboratively.
License¶
MIT License — Fully open source, free for both academic and commercial use.