Event
Tsinghua AIR proposes OxyGen unified KV Cache management
1 reports · 1 sources
The Institute for AI Industry Research (AIR) at Tsinghua University, together with the Department of Electronic Engineering at Tsinghua University and the University of Science and Technology of China, proposes OxyGen, which through unified KV Cache management enables embodied models' action and language reasoning to share prefix computation, thereby allowing robots to handle action and language tasks simultaneously. The related paper has been accepted by NeurIPS 2026, and the method has been validated and open-sourced on embodied reasoning frameworks and models such as openpi, StarVLA, and Xiaomi-Robotics-0. According to the team, when evaluating π₀.₅ on an RTX 4090, action generation frequency and total language throughput simultaneously reached up to 3.7 times those of independent execution; on the onboard Jetson AGX Thor of a humanoid robot, the total time for action and language reasoning dropped from 1030 milliseconds to 393 milliseconds.
Report timeline
- Tsinghua AIR proposes OxyGen: unified KV Cache management lets robots talk while acting
The Institute for AI Industry Research (AIR) at Tsinghua University, together with Tsinghua's Department of Electronic Engineering and the University of Science and Technology of China, has proposed OxyGen, which uses unified KV Cache management to let an embodied model's action and language reasoning share prefix computation. The paper has been accepted to NeurIPS 2026.
CVer · 微信公众号 ·