π0.5 · actions & language
faster in both outputs
action frequency (H=10)
language throughput
Speedup across backbones
Xiaomi-Robotics-0
StarVLA
§ 1 · Demo
Same task. Less waiting.
Replay data could not be loaded. Watch the annotated video.
The baseline finishes each language response before moving. OxyGen carries language generation across action updates.
Both runs use the same frozen π0.5 policy and task; OxyGen changes the inference schedule and KV-cache handling, not the policy or target actions. Action inference, language inference, and execution are shown as separate phases reconstructed from the logged run.
§ 2 · Method
One cache. Across tasks and time.
Share the observation.
Batch language across frames.
The illustration uses N = 12 tokens per request and k = 4 decoding steps per frame. After warm-up, 3 requests share a batch. A check marks completion and cache eviction. Each request keeps its own observation prefix and generated-token KV state.
H = 10 actions, S = 10 denoising steps, N = 30 tokens per language request, k = 5 decoding steps per frame. At steady state, six unfinished requests share a batch.
Totals are derived from measured action frequency in the paper’s ablation (10 / Hz). Prefill ≈ 44.0 ms and denoising ≈ 20.4 ms come from a separate synchronized camera-ready profile. Stage widths are illustrative; the remainder includes decoding and other frame work.
§ 3 · Results
Faster actions and more language.
Measurements on desktop and on-board GPUs, followed by transfer across backbones and model sizes.
| Configuration | k = 1 / k = 5 · 0–8× |
|---|---|
| Across architecturesSharing and batching transfer beyond π0.5. | |
| π0.5 | 3.71× / 2.98× |
| Xiaomi-Robotics-0 | 5.83× / 3.23× |
| StarVLAQwen3-VL-4B + PI_v3 | 6.64× / 3.60× |
| More expertsSpeedup holds as experts increase. | |
| + GR00T3 experts | 6.05× / 3.53× |
| + GR00T + OFT4 experts | 6.30× / 3.69× |
| Across backbone sizesThe gain persists as the backbone grows. | |
| Qwen3.5-0.8B | 6.30× / 3.58× |
| Qwen3.5-2B | 6.31× / 3.62× |
| Qwen3.5-4B | 6.55× / 3.71× |
| Qwen3.5-9B | 7.18× / 3.79× |
Shared details for the throughput, batching and transfer experiments.
The throughput and batching panels use π0.5 on LIBERO; the transfer table reports each model’s 30-token row against its own isolated baseline. k is the per-frame decoding budget; S = 10 for π0.5. The throughput chart compares RTX 4090 and Jetson AGX Thor, while the batching chart uses k = 5. Ringed points match the π0.5 workload.
Transfer runs use one RTX 4090 and an i9-13900K (Xiaomi S = 5; StarVLA/Qwen3.5 S = 4). Runtime scaling uses random parameters; policy quality is validated separately with the official checkpoint.
10 seeds, 20,000 rollouts, official π0.5-LIBERO checkpoint. OxyGen changes the inference schedule only; the policy and action outputs are unchanged.
RTX 4090, OxyGen batch size 4. The table compares isolated execution with batched decoding, which reads the VLM weights once for several requests. Parallel processes (MPS) use 12.49 GiB and 120.9 J per request.
§ 4 · On-board measurement
Lower latency, on board.
2.6× faster inference on the on-board Jetson AGX Thor, with room for language during motion.
π0.5 on Jetson AGX Thor: H = 10, S = 10, N = 30, k = 5. Input: three 224 × 224 RGB views. Action timings include prefill and denoising; the speedup compares total inference time.
The motion window illustrates where physical execution could overlap language after action dispatch. It uses measured inference stages and a 333 ms execution window; communication overhead is omitted. All bars share the same time scale.
§ 5 · Resources
Read it, run it, build on it.
@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}
}