Paper
3we: An Open Infrastructure for Sim-to-Real Embodied AI Research
Abstract
Embodied AI research requires tight integration of simulation, real hardware, and learning algorithms — yet existing platforms force researchers to choose between expensive proprietary systems, simulation-only environments, or hardware platforms with steep ROS2 learning curves. We present 3we, a fully open-source infrastructure where the same Python code runs identically across a lightweight mock simulator, Gazebo, NVIDIA Isaac Sim, and real hardware (Raspberry Pi 5 + Hailo-8L) with zero modification. The platform provides an AI-First Python API that reduces typical ROS2 navigation code from 60+ lines to 5, Gymnasium-compatible environments, native VLM/VLA integration, and standardized benchmarks across 8 evaluation scenes. On reproducible hardware costing under $300 (BOM), preliminary simulation results indicate Sim-to-Real transfer ratios targeting 0.6 on point navigation tasks, with simulated baselines achieving 84.3% success rate (SPL 0.72) in structured environments.
BibTeX Citation