Vision-Language-Action (VLA) models have recently achieved promising performance in robotic manipulation. However, existing benchmarks mainly evaluate generalization on static manipulation tasks and largely overlook dynamic interaction scenarios. To address this gap, we present ReflexBench, a benchmark for reaction-critical manipulation. ReflexBench contains six dynamic tasks and introduces an evaluation framework that decouples simulator stepping from robot control while supporting configurable latency under synchronous and asynchronous inference. Building upon ReflexBench, we propose ReflexVLA, an efficient VLA model designed for reaction-critical manipulation without large-scale robot-data pretraining. ReflexVLA enhances temporal reasoning through latent future prediction and multi-frame temporal fusion within the vision backbone, while reducing deployment latency through batched visual encoding and CUDA Graph replay. Experiments show that ReflexVLA consistently improves dynamic manipulation performance while maintaining competitive accuracy on standard static manipulation benchmarks, and real-world experiments further demonstrate its effectiveness under practical deployment conditions. Project website: https://reflexvla.github.io
Figures & tables
Fig. 1 : Overview of our work. We introduce ReflexBench , a benchmark of six reaction-critical manipulation tasks. Building upon it, we propose ReflexVLA , a fast and predictive VLA that achieves strong performance with low latency.
Fig. 2 : Illustration of Four Inference Mechanisms. ReflexBench models the effect of policy inference latency by explicitly simulating the temporal gap between perception and action execution while allowing the latency to be adjusted in a controllable manner.
Fig. 3 : Overview of ReflexVLA. ReflexVLA is an efficient vision-language-action model for reaction-critical robotic manipulation, integrating latent future prediction, multi-frame temporal fusion, and inference latency optimization to enable fast and anticipatory decision-making.
Fig. 4 : Success rates of SmolVLA under synchronous and asynchronous inference across different inference frequencies on two tasks. Chunk size is set to 8.
Fig. 5 : Success rate matrices of SmolVLA under different chunk sizes and action horizons on two tasks. Chunk size is the policy’s training hyperparameter, while action horizon is the number of action steps executed during inference. The inference frequency is fixed at 30 Hz.
Model
Params
Conveyor Belt Pick-and-Place
Ball Catching
Whack-a- Mole
Rolling Ball Interception
Ball Throwing
Rotating Peg Insertion
Avg.
OpenVLA-OFT [ 20 ]
7B
58.0 ± 1.2
5.3 ± 1.2
100.0 ± 0.0
41.4 ± 4.2
10.0 ± 3.3
1.3 ± 0.7
36.0
π0.5 [ 21 ]
4B
39.1 ± 1.0
6.0 ± 0.7
98.9 ± 0.3
36.8 ± 1.8
34.0 ± 2.7
6.7 ± 1.4
36.9
PUMA [ 5 ]
4B
67.4 ± 1.7
4.0 ± 0.7
100.0 ± 0.0
85.1 ± 3.4
33.8 ± 1.0
11.1 ± 1.7
50.2
DynamicVLA [ 4 ]
0.5B
30.6 ± 2.4
4.4 ± 1.0
90.2 ± 1.7
45.8 ± 1.7
36.2 ± 4.7
16.7 ± 4.2
37.3
SmolVLA [ 8 ]
0.5B
19.3 ± 2.0
2.9 ± 0.8
100.0 ± 0.0
70.2 ± 2.0
27.1 ± 1.0
10.0 ± 4.6
38.3
VLA-Adapter [ 10 ] (Baseline)
1B
36.8 ± 2.4
6.0 ± 2.0
68.4 ± 3.9
23.1 ± 7.7
29.1 ± 5.2
18.4 ± 2.5
30.3
TABLE I : Performance comparison on ReflexBench. Results are reported as mean success rate (%) ± standard deviation over three runs.
Fig. 6 : Real-world rollout examples. From top to bottom: Conveyor Belt Pick-and-Place, PressButtons, and CatchBalls.
Model
Spatial
Object
Goal
Long
Avg.
OpenVLA-OFT [ 20 ]
97.6
98.4
97.9
94.5
97.1
π0.5 [ 21 ]
98.8
98.2
98.0
92.4
96.9
VLA-Adapter [ 10 ]
97.8
99.2
97.2
95.0
97.3
ReflexVLA (Ours)
98.2
99.2
98.0
93.6
97.2
TABLE II : Results on LIBERO. Accuracy is reported in (%).
Method
Variant
SR (%)
Lat. (ms)
Baseline
–
36.8
81.522
+ Future Pred.
Trainable
4.9 (-31.9)
82.508
Frozen
62.8 (+26.0)
82.508
+ Temporal Fusion
Cross Attn.
66.1 (+29.3)
127.335
MHA
68.2 (+31.4)
125.824
MHA (Middle)
71.7 (+34.9)
125.107
TABLE III: Progressive ablation study of ReflexVLA.
Model
Conveyor
PressButtons
CatchBalls
SmolVLA [ 8 ]
2/20
0.9
3.8
PUMA [ 5 ]
13/20
20.8
5.4
Ours
16/20
22.5
6.7
TABLE IV: Quantitative results of the real-world experiments. Conveyor Belt reports the number of successful trials out of 20 attempts. PressButtons reports the average number of buttons pressed within 30 seconds, and CatchBalls reports the average number of balls caught out of 10 balls. Each policy is evaluated 20 times on every task.
The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China. · Huazhong University of Science and Technology, Wuhan, China. · Westlake University, Hangzhou, China. +1