Reflex: Enabling Fast and Predictive Vision-Language-Action Models for Reaction-Critical Manipulation

Yuxuan Chen1, Wanruo Zhang1, Xiao Li1
1Shanghai Jiao Tong University
CodeAfter acceptance Paper arXiv

Abstract

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.

Overview

Challenges, ReflexVLA solution, and the six ReflexBench tasks
Overview of Reflex. ReflexBench evaluates six reaction-critical manipulation tasks, while ReflexVLA combines future prediction, temporal context, and low-latency inference.

ReflexBench

ReflexBench is a latency-aware benchmark for manipulation tasks in which objects and goals continue to evolve while the policy is computing. It contains six tasks: Conveyor Belt Pick-and-Place, Ball Catching, Whack-a-Mole, Rolling Ball Interception, Ball Throwing, and Rotating Peg Insertion.

Unlike benchmarks that pause the environment during policy inference, ReflexBench decouples simulation from robot control. It supports both synchronous and asynchronous inference and injects configurable latency, enabling controlled evaluation of perception-execution delay under practical deployment conditions.

Synchronous and asynchronous inference mechanisms in ReflexBench
Latency-aware evaluation. ReflexBench reproduces the temporal gap between observation, policy inference, and action execution for synchronous and asynchronous control.

ReflexVLA

ReflexVLA is a compact 1B-parameter VLA designed to react quickly while reasoning about future scene dynamics. It combines three complementary components:

  • Latent future prediction learns anticipatory representations by predicting future visual features from a frozen DINOv3 encoder.
  • Multi-frame temporal fusion integrates short-term motion cues inside the vision backbone without increasing the number of visual tokens consumed by the language model.
  • Inference latency optimization uses batched visual encoding and CUDA Graph replay to reduce deployment overhead.
ReflexVLA architecture
ReflexVLA architecture. Temporal fusion, latent future prediction, and system-level acceleration enable fast, anticipatory control.

Experiments

Latency-Aware Inference

Success rates under different inference paradigms and frequencies
Inference frequency. High-frequency asynchronous inference improves performance by enabling continuous action execution.
Success rate matrices for action chunk size and horizon
Chunking. Larger chunks with shorter action horizons perform best under high-frequency asynchronous inference.

Main Results on ReflexBench

Performance comparison on ReflexBench. Results are mean success rate (%) ± standard deviation over three runs.

Model Params Conveyor Belt
Pick-and-Place		 Ball
Catching		 Whack-a-
Mole		 Rolling Ball
Interception		 Ball
Throwing		 Rotating Peg
Insertion		 Avg.
OpenVLA-OFT7B58.0 ± 1.25.3 ± 1.2100.0 ± 0.041.4 ± 4.210.0 ± 3.31.3 ± 0.736.0
π0.54B39.1 ± 1.06.0 ± 0.798.9 ± 0.336.8 ± 1.834.0 ± 2.76.7 ± 1.436.9
PUMA4B67.4 ± 1.74.0 ± 0.7100.0 ± 0.085.1 ± 3.433.8 ± 1.011.1 ± 1.750.2
DynamicVLA0.5B30.6 ± 2.44.4 ± 1.090.2 ± 1.745.8 ± 1.736.2 ± 4.716.7 ± 4.237.3
SmolVLA0.5B19.3 ± 2.02.9 ± 0.8100.0 ± 0.070.2 ± 2.027.1 ± 1.010.0 ± 4.638.3
VLA-Adapter (Baseline)1B36.8 ± 2.46.0 ± 2.068.4 ± 3.923.1 ± 7.729.1 ± 5.218.4 ± 2.530.3
ReflexVLA (Ours)1B73.8 ± 4.17.3 ± 0.7100.0 ± 0.077.1 ± 7.131.7 ± 1.012.4 ± 1.550.4

ReflexVLA achieves the highest average success rate on ReflexBench. It improves the VLA-Adapter backbone from 30.3% to 50.4%, outperforms substantially larger general-purpose VLAs, and matches PUMA while using only one quarter of its parameters.

50.4%Average ReflexBench success
1BModel parameters
65.0 msOptimized inference latency
97.2%Average LIBERO success

Results on LIBERO

Success rate (%) on the standard static manipulation benchmark.

ModelSpatialObjectGoalLongAvg.
OpenVLA-OFT97.698.497.994.597.1
π0.598.898.298.092.496.9
VLA-Adapter97.899.297.295.097.3
ReflexVLA (Ours)98.299.298.093.697.2

Although ReflexVLA is designed for dynamic manipulation, it maintains a 97.2% average success rate on LIBERO, comparable to the best-performing VLAs.

Ablation Study

Progressive ablation of ReflexVLA. SR denotes success rate and Lat. denotes inference latency.

MethodVariantSR (%)Lat. (ms)
Baseline—36.881.522
+ Future Pred.Trainable4.9 (−31.9)82.508
Frozen62.8 (+26.0)82.508
+ Temporal FusionCross Attn.66.1 (+29.3)127.335
MHA68.2 (+31.4)125.824
MHA (Middle)71.7 (+34.9)125.107
+ Latency Opt.Batch Enc.
+ CUDA Graph		73.8 (+37.0)64.991

Frozen latent targets provide stable predictive supervision, temporal fusion at intermediate visual features best captures motion, and batched encoding with CUDA Graph replay reduces latency from 125.1 ms to 65.0 ms while further improving success.

Real-World Experiments

Results over 20 evaluations per task. Conveyor reports successful trials out of 20; PressButtons reports buttons pressed within 30 seconds; CatchBalls reports balls caught out of 10.

ModelConveyorPressButtonsCatchBalls
SmolVLA2/200.93.8
PUMA13/2020.85.4
ReflexVLA (Ours)16/2022.56.7
Real-world ReflexVLA rollout examples
Real-world rollout examples. From top to bottom: Conveyor Belt Pick-and-Place, PressButtons, and CatchBalls.

Citation

@article{chen2026reflex,
  title         = {Reflex: Enabling Fast and Predictive Vision-Language-Action
                   Models for Reaction-Critical Manipulation},
  author        = {Chen, Yuxuan and Zhang, Wanruo and Li, Xiao},
  journal       = {arXiv preprint arXiv:2608.14379},
  year          = {2026},
  eprint        = {2608.14379},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2608.14379}
}