On September 15, Infinigence AI, in partnership with Tsinghua University and Shanghai Jiao Tong University, open-sourced APXInf, an edge inference engine for embodied intelligence, hosted on GitHub at RLInf/APXInf-robo. The project's goal is concrete: to make embodied models run on real robots not just "at all," but fast, stably and easily enough to be practical under real compute, power and cost constraints.
The headline number is self-reported: π0.5 FP8 end-to-end inference latency on Jetson AGX Thor dropped from 278ms to under 26ms, roughly 38.46Hz, meeting the response requirements of real-time robot control. A 10.7x cut in latency compresses the perceive-decide-act loop that governs robot behavior. A caveat is in order: these figures come from the project team, not from independent evaluation. Community testing (Datawhale, September 17) re-measured π0.5 on an RTX 4090 at 31.38ms per step in BF16 and 25.99ms in INT8 — close in spirit to the official numbers but on different hardware, so not directly comparable.
The engineering approach is worth spelling out. APXInf assigns memory allocation and operator execution to explicit static paths per model, captures the compute flow with CUDA Graphs, and reuses pre-allocated buffers to cut repeated overhead. A minimal runtime is built in Rust for memory safety and auditability, with Python interfaces on top so developers keep their usual workflow. The more distinctive choice is Agentic Engineering: expert experience in model porting, performance tuning and deployment validation is packaged into workflows and skills that coding agents can execute (for example, skills/model-port-workflow), letting AI participate in onboarding and optimizing new models. Most inference frameworks still assume manual adaptation by human engineers, which is why the project describes itself as "agentic native."
On task performance, in the official LIBERO-10 evaluation π0.5 scored 92.2% (FP8) and 92.8% (BF16) on Jetson AGX Thor, versus 92.4% for the reference implementation — speedup without a measurable hit to task success. APXInf currently supports two embodied models, π0.5 and WALL-OSS, across Jetson Orin, Jetson Thor and RTX 4090. The roadmap lists Qwen and Groot support, extreme nvfp4 optimization, plus domestic chip backends, domestic robot operating systems and AMD.
The context matters. In September 2025, Infinigence AI and Tsinghua open-sourced RLinf, described as the world's first large-scale reinforcement-learning training framework for embodied intelligence. APXInf is the first edge-inference project in that ecosystem, linking model training, evaluation and deployment into one chain.
The engineering consensus in embodied AI is that cloud chatbots can tolerate hundreds of milliseconds of latency; robots cannot. Under continuous perception and real-time control, a few hundred milliseconds means lagging motion, broken trajectories and failed tasks. APXInf targets exactly this last mile that large-scale adoption cannot bypass.
The boundaries are equally clear. Open source and SOTA numbers are not proof of large-scale deployment. Long-term stability on real robots, tolerance of latency jitter and memory pressure, and portability across bodies are questions only the physical world can answer. Whether this engine really "fills the missing link" is a verdict the developer community will deliver through actual deployments.
There is a broader pattern here worth noting. As embodied models get better at understanding scenes and instructions, the bottleneck shifts from model quality to the runtime that carries perception into action. Edge inference is where most of the practical engineering now lives: small batch sizes, multi-view inputs, strict latency budgets and tight memory. Open-sourcing a tuned runtime rather than keeping it proprietary is itself a statement — the project's sponsors are betting that the ecosystem, not the moat, drives adoption. That logic worked for training frameworks; whether it transfers to inference remains to be seen, but it explains why APXInf ships with porting workflows for agents rather than a closed benchmark suite.
[1][2]