Infinigence AI, Tsinghua University and Shanghai Jiao Tong University release APXInf embodied edge inference engine

Infinigence AI announced on September 15, 2026, that it had officially open-sourced APXInf, an edge inference engine for embodied intelligence, in collaboration with Tsinghua University and Shanghai Jiao Tong University. The engine supports RTX 4090, Jetson Orin and Jetson Thor, and is designed to provide an end-to-end workflow from model development to deployment on robots, the company said. According to the announcement, APXInf cut end-to-end inference latency for PI 0.5 FP8 from 278 ms to less than 26 ms, reaching an inference frequency of 38.46 Hz.

Infinigence AI is presenting an edge-side design to address inference latency, a common bottleneck in putting robots to practical use. The reduction from 278 ms to less than 26 ms and the 38.46 Hz rate cited in the announcement show how the choice of computing platform can affect product viability in applications requiring real-time control.

What is new in this announcement is that it treats embodied intelligence not as a matter of “model performance,” but as an inference system running on the robot itself, and open-sources an edge inference engine. The focus on low latency, low power consumption and high stability extends an established approach; these have long been unavoidable requirements for real-time robot control. APXInf’s distinguishing feature is that it addresses those challenges across multiple computing platforms: RTX 4090, Jetson Orin and Jetson Thor. In the context of this newspaper’s previous coverage, leaderobot.com reports remote operation of a Shanghai robot by a Beijing patient with a spinal cord injury showed how, in situations involving connections to physical robots, such as remote operation, communication and control latency directly affect the user experience and movement. APXInf, likewise focused on real-world robot operation, puts the emphasis on reducing latency by running inference on the robot rather than in the cloud. In other words, while the earlier report depicted the “remote connection” phase, this announcement concerns building the infrastructure for the “on-robot” phase. The industry-wide implication is that competition in embodied intelligence is shifting beyond model accuracy to the placement and optimization of inference systems. APXInf demonstrates that responsiveness can vary significantly depending on the edge execution environment, even when the model is the same. For robotics companies, where to run a trained model and which computing resources to rely on become constraints on commercialization. If software optimization is not aligned with the implementation requirements of Jetson and GPU platforms, deployment on physical robots could be delayed. Questions remain about how reproducible APXInf’s performance is across different devices, what workload conditions underpin the 38.46 Hz rate, and how many external developers will integrate it into physical robots after its release as open source. It is also not yet clear how far the implementation published on GitHub will be standardized in mass-produced robots.

The open-source release of an inference system designed for edge computing platforms such as Jetson Thor and Jetson Orin underscores the importance of hardware-level optimization for Japanese robotics companies and embedded software developers. Companies working on low-latency control or physical-robot integration, in particular, need to focus not only on model accuracy but also on compatibility with the execution platform.