3we: Open-source AI-first Python API for robot navigation with $300 hardware
Switch between simulation and real hardware with one line of code change.
3we addresses the steep learning curve of ROS2 for embodied AI researchers by offering a clean Python API that abstracts away launch files, topics, and services. Users write simple async Python code to get camera images, LiDAR scans, and navigate using Nav2 under the hood. The platform supports four interchangeable backends: mock (zero-dependency 2D kinematics), Gazebo Harmonic, NVIDIA Isaac Sim, and real hardware via ROS2. This allows seamless simulation-to-real transfer with minimal code changes. Benchmarks include 7 standardized scenes compatible with Gymnasium for RL training.
The reference hardware is fully open-source under CERN-OHL-P v2 and costs only $300 in BOM. It uses a Raspberry Pi 5, Hailo-8L AI accelerator, ESP32-S3 with micro-ROS, LD06 LiDAR, and Mecanum wheels. A standout feature is VLM-controlled navigation: researchers can simply instruct the robot in natural language (e.g., "find the red bottle") using GPT-4o, Qwen-VL, or local LLaVA. The system runs a perception-action loop internally. 3we is available on GitHub (Apache 2.0) and PyPI, with full documentation and a paper.
- Four backends (mock, Gazebo, Isaac Sim, real) with identical Python API, enabling trivial sim-to-real transfer.
- $300 open-hardware BOM including Raspberry Pi 5, Hailo-8L, LD06 LiDAR, and Mecanum wheels.
- VLM-controlled navigation: use GPT-4o or local models to command robots via natural language.
Why It Matters
Lowers barriers for embodied AI research by making ROS2 accessible and providing affordable open hardware for real-world deployment.