SR-VLA: Stall-Reduced VLA Execution via Predicted-Input Lookahead and Selective KV Cache Repair
Authors: Tao Zhang, Yi Zhong, Jinhao Ruan, Yipeng Gao, Linfeng Zhong, Youming Yang, Yuan Wang
Organizations: School of Integrated Circuits, Peking University · Key Laboratory of Microelectronic Devices and Circuits, MPW Center, Peking University · Beijing Advanced Innovation Center for Integrated Circuits
Vision-language-action (VLA) policies incur control stalls when inference blocks the next action chunk. Asynchronous early inference can hide latency, but its results may be inconsistent with the arriving actual inputs. Unconditional reuse risks acting on mismatched inputs, whereas full recomputation discards useful precomputation. We propose SR-VLA, combining predicted-input lookahead with selective KV cache repair. During current chunk execution, a lightweight action-conditioned predictor generates next-boundary visual observations and proprioceptive state. These predictions, together with the task instruction, drive the frozen VLA to precompute a speculative KV cache and action chunk. After real inputs are available to VLA, predicted--real differences and task conditions determine whether to accept the action chunk or trigger selective KV repair. On LIBERO, SR-VLA achieves 2.67× and 1.27× query speedups over the respective synchronous π0.5 and OpenVLA-OFT baselines, while maintaining competitive average task success rates. Under target-motion perturbation, success remains baseline-comparable, with 1.52×--1.93× query speedup.
Figures & tables
Fig. 1: Motivation and design principle of SR-VLA. (a) Synchronous inference exposes the full next-decision inference latency. (b) An asynchronous schedule using launch-time images illustrates the timing trade-off. The right panel summarizes the remaining gap and SR-VLA principle. Here, K denotes the number of action steps executed between consecutive VLA decisions, so t+K is the next action-chunk boundary.
Fig. 2: SR-VLA execution path. During committed At , predicted images and proprioceptive state drive the frozen VLA to construct a speculative KV Cache and next action. At the target boundary, real inputs, task relevance, and protection conditions selectively repair KV Cache.
Fig. 3: Joint next-boundary image and proprioceptive-state prediction, conditioned on the current image, robot state, and committed action chunk.
Setting
Patch cosine similarity ↑
Proprio. MSE ↓
Closed-loop success (%) ↑
τd=0.5
τd=0.95
τd=1.0
Sync ( π0.5 , real input)
1.000
1.000
1.000
0.000
97.8
Async (w/o prediction)
-0.037
0.394
0.856
0.069
45.2
Predictor
0.793
0.854
0.963
0.004
87.6
Ablation variants
w/o FiLM
0.551
0.697
0.924
0.004
70.6
TABLE I: Next-boundary predictor ablation on LIBERO-Spatial with frozen π0.5 ; boundary repair is disabled.
Fig. 4: Selective KV repair schematic. (a) predict-real error and task relevance on image patches. (b) Visual KV replacement.
Repair setting
Task success (%) ↑
Action dev. ↓
Repair rate (%)
Updated KV (all-query avg., %)
Full policy
97.8
0.166
31.68
27.37
w/o input-driven gate
91.4
0.211
20.2
17.1
w/o task-protection gate
94.6
0.176
30.8
25.4
w/o Top- k error
91.0
0.246
36.3
0.5
w/o Top-ratio risk
95.4
0.163
34.1
27.6
TABLE II: Two-level repair-policy ablation with the predictor and frozen π0.5 backbone fixed at K=5 .
Fig. 5: Latency and target-motion perturbation results. (a) Normal-load mean stall. (b) Mean stall under controlled delay. (a)–(b) separate exposed prediction from forward computation with the real inputs. (c) LIBERO-Spatial task success and SR-VLA repair rate under target displacement, where 0 cm denotes no motion.
Method
Success (%) ↑ S/O/G/L Avg.
Query/ Action (ms) ↓
Stall Mean (ms) ↓
Residual Stall ( × ) ↓
Speedup ( × ) ↑
π0.5 -Sync [ 1 ]
97.8 /97.4/97.0/93.4
96.4
25.62
128.11
1.00
1.00
OpenVLA-OFT-Sync [ 4 ]
97.6/ 98.4 /97.9/94.5
97.1
16.90
135.18
1.06
1.52
VLASH ( π0.5 ) [ 14 ]
97.2/ 99.6 /95.6/ 94.0
96.6
12.05
60.25
0.47
2.13
VLA-Cache (OFT) [ 19 ]
98.3/97.5/ 98.3 / 95.4
97.4
15.24
121.88
0.95
1.68
SR-VLA ( π0.5 )
97.8 /99.2/ 97.2 /93.4
96.9
9.59
47.95
0.37
2.67
SR-VLA (OFT)
98.8 /98.2/97.6/94.2
97.2
13.27
106.19
0.83
1.93
TABLE III: Task comparisons on LIBERO.
Suite
τimg
τp
Cmax
Error ke (per view)
Risk αr (ratio)
Repair rate (%)
Updated KV (all-q., %)
Spatial
0.58
0.40
27
220
0.015
31.68
27.37
Object
1.07
0.59
17
147
0.049
43.27
24.88
Goal
0.57
0.47
29
137
0.018
30.09
16.24
Long
0.30
0.58
15
88
0.008
38.88
13.44
TABLE IV: Suite-wise repair settings and specifications.