AsyncVLA: Asynchronous Flow Matching for Vision-Language-Action Models
Fuzhou University AsyncVLA : Asynchronous Flow Matching for Vision-Language-Action Models : A rXiv preprint : 2025.11 : Yuhua Jiang, Shuang Cheng, Yan Ding, Feifei Gao, Biqing Qi : , AI 20 26 - 4-1 AsyncVLA
1.摘要
1. VLA FM VLA FM SFM SFM VLA AsyncVLA FM AFM AsyncVLA token VLA SFM 使 SFM AFM KV 广 AsyncVLA AsyncVLA AFM
2.引言
2. VLA VLA FM VLA token FM SFM SFM
2.引言
2. token VLA
2.引言
2. SFM VLA : 1 VLA AsyncVLA VLA SFM FM AFM AFM SFM 2 SFM token AsyncVLA 使 访 3 AsyncVLA VLA
3.相关工作
3. (1)VLA VLA 使 VLM VLA VLM CoT CoT VLA FlowVLA CoT OpenVLA OFT π0 π0.5 WALL-OSS EO-1 FM SFM SFM AsyncVLA AFM
3.相关工作
3. (2) VLA VLA CollabVLA ReflectVLM SC-VLA RB-VLA 使 RL SFT DLLM VLA LLaDA VLA dVLA UD-VLA CoT KV token AsyncVLA SFM AFM
4.模型-概述
4. - SFM ---> Confidence Rater SFM ( 0/ 1) ---> AFM 0 Context 1 FM
4.模型-1
4. -1 Asynchronous Flow Matching VLM FM VLA L Velocity FM ot: t l: τ V VLM
4.模型-1
4. -1 Asynchronous Flow Matching AFM SFM m∈RL m 1 0 AFM 0 1 使 Euler
4.模型-1
4. -1 Time Embedding Transformer token VLM d 1. FM τ s · τ m 2. 沿 S τ m 线 P ˆ a τ t:t+L RL×d 3. MLP h τ t:t+ ˆ x τ t:t+L∈RL×d VLM ˆ x τ t:t+L∈RL×d VLM transformer 4. VLM Transformer Full Attention token
4.模型-1
4. -1 SFM AFM 使 SFM AFM VLM FM token AFM ut:t+L a τ t:t+L
4.模型-1
4. -1 SFM AFM
4.模型-2
4. -2 Confidence Rater SFM SFM MSE et:t+L α β ε 0 max{el} min{el} MSE α=0.01 β=0.98 ε=1×10−6 pl:t+l qt:t+L MSE Loss
4.模型-2
4. -2 Confidence Rater AsyncVLA actiontoken token VL SFM L
4.模型-2
4. -2 Confidence Rater transformer 线 308 M VLA 4.08 B 7.56% 使 线 token VL token transformer full attention VL token 使 Sigmoid pl 使 pl l 1{·} T∈ 0,1 token 0.5
5.实验
5. Qwen2.5-VL-3BInstruct VLM FM AsyncVLA Open X-Implementation LIBERO Bridge-V2 Fractal 4 H200 GPU 8 GPU 8 GPU H200GPU
5.实验
5. Libero
5.实验
5. Bridge- V2 WidowX Fractal Google Robot
5.消融实验
5. WidowX w/o Unified Training: 2 SFM AFM SFM w/o AFM Inference: AFM 使 SFM w/o Confidence Rater: AFM 0.5
6.结论
6. AsyncVLA AsyncVLA AFM 使 AsyncVLA token SFM AFM KV - cache AsyncVLA token 广 AsyncVLA
谢谢!