APXInf: Open-Source Edge Inference Engine for Robotics

APXInf, an open-source edge inference engine for embodied AI, reduces reaction time from 278ms to under 26ms on Jetson Thor for the PI 0.5 model, a 10.7x improvement. It supports PI 0.5 and WALL-OSS models on RTX 4090, Jetson Orin, and Thor.
Key insights
Performance Improvement
APXInf reduces reaction time from 278ms to under 26ms for PI 0.5 on Jetson Thor, a 10.7x improvement. This enables real-time control for robotics. The engine is designed for long-running stability without jitter or disconnection.
Compatibility and Ecosystem
APXInf supports PI 0.5 and WALL-OSS embodied models. It runs on desktop RTX 4090 and robot platforms Jetson Orin and Thor. It is part of the same ecosystem as the open-source training framework RLinf, enabling a unified toolchain from training to deployment.