π₀.₅
Paper · LIBERO · S=10
OxyGen
Unified KV Cache Management for VLA Inference under Multi-Task Parallelism
TL;DR OxyGen treats the KV cache as a shared resource across experts and control frames, enabling action and language generation to run together without redundant context encoding or blocked control.
1 Institute for AI Industry Research (AIR), Tsinghua University 2 Department of Electronic Engineering, Tsinghua University 3 University of Science and Technology of China
LIBERO demo
Concurrent action and
textual-memory generation
Measured model inference and 1× simulator timing. OxyGen advances unfinished text requests across frames while the isolated openpi baseline blocks on each request.
Measured rollout
model inference + 1× simulator time Put the white mug on the plate and the chocolate pudding to the right of the plate
OxyGen vs. isolated openpi baselineπ0.5 · RTX 4090 & Jetson AGX Thor
3.7×up to end-to-end speedup
>200language tokens / second
70 Hzaction frequency, simultaneously
Challenge and method
From isolated execution
to unified KV management.
Isolated action and language paths repeat the VLM prefix and compete for the same device. OxyGen shares the prefix across experts and batches unfinished language requests across control frames.
Paper evaluation
Higher throughput without
trading away control.
Evaluated with π0.5 on RTX 4090 and Jetson AGX Thor across LIBERO, DROID, and ALOHA workloads, including an on-board humanoid deployment.
End-to-end tradeoffMain result
Ablationπ0.5-LIBERO
Extended validation
The same system idea
transfers to other MoT VLAs.
The π0.5 curve is from the paper evaluation; Xiaomi-Robotics-0 and StarVLA are later validations using their native implementations on the same RTX 4090. Here, N is the language decoding length and k the decoding steps per request in each frame.
Xiaomi-Robotics-0
4.7B · native implementation · S=5
StarVLA Qwen3VL-PI_v3
5.07B · native implementation · S=4
More heterogeneous experts
StarVLA · shared Qwen3-VL-4B prefix
Model-size scaling
StarVLA Qwen3.5 backbones
StarVLA’s heterogeneous heads were trained separately; these measurements validate system scaling and shared-prefix fan-out, not the quality of a jointly trained multi-expert policy.
LIBERO textual memory
Action and observation-grounded
memory, generated together.
We adapt π0.5 with suffix-only LoRA to describe completed task progress from the current observation. The action path keeps the base model and shared prefix unchanged.
338,575annotated frames
87.25%held-out exact match
8.16 msmedian 11-token LoRA overhead
0action difference with adapter zeroed
More rollouts
Videos load only when played.
Alphabet soup & tomato sauce
Place both objects in the basket.
Two moka pots
Place both moka pots on the stove.
Open resources
Reproduce the system.
Extend the example.
Code, checkpoints, annotations, and the paper are available as separate entry points so each part can be reused independently.
PaperMethod and evaluation
CodeJAX and PyTorch backends
Adapterπ0.5 textual-memory LoRA
DatasetLIBERO textual-memory annotations
Citation
If OxyGen supports your work, please cite:
@article{li2026oxygen,
title={OxyGen: Unified KV Cache Management for VLA Inference under Multi-Task Parallelism},
author={Li, Xiangyu and Tang, Huaizhi and Ding, Xin and Wang, Weijun and Cao, Ting and Liu, Yunxin},
journal={arXiv preprint arXiv:2603.14371},
year={2026}
}