StarVLA-$α$: Reducing Complexity in Vision-Language-Action Systems

Jinhui Ye; Ning Gao; Senqiao Yang; Jinliang Zheng; Zixuan Wang; Yuxin Chen; Pengguang Chen; Yilun Chen; Shu Liu; Jiaya Jia

Source-linked AI summary

StarVLA-$α$: Reducing Complexity in Vision-Language-Action Systems

Jinhui Ye, Ning Gao, Senqiao Yang, Jinliang Zheng, Zixuan Wang, Yuxin Chen, Pengguang Chen, Yilun Chen, Shu Liu, Jiaya Jia

arXiv:2604.11757v1Apr 13, 2026cs.ROcs.AIcs.CV

TL;DR

VLA systems are difficult to compare because their architectures, data, embodiments, and benchmark engineering vary substantially. StarVLA-α addresses this with a simple VLM-based baseline and controlled analyses, finding strong multi-benchmark performance and supporting a streamlined, reproducible approach.

Problem
VLA research lacks clear evidence about which components drive performance because architectures, pretraining data, embodiments, and benchmark-specific engineering vary substantially.
Method
StarVLA-α uses a strong pretrained Qwen3-VL backbone, a lightweight action head, minimal data processing, and controlled comparisons across key VLA design choices.
Results
Across unified training on four benchmarks, the simple baseline remains highly competitive; it achieves 98.8% average success on LIBERO and +6.8% on SimplerEnv Google VM.
Takeaways & Limitations
A strong VLM with a straightforward action head and minimal preprocessing can provide a competitive, reproducible baseline for generalist VLA research.
Takeaways & Limitations
The framework relies on a minimal-sufficiency hypothesis that a strong VLM and lightweight action head capture most benefits attributed to complex designs.

Abstract

from arXiv · show

Vision-Language-Action (VLA) models have recently emerged as a promising paradigm for building general-purpose robotic agents. However, the VLA landscape remains highly fragmented and complex: as existing approaches vary substantially in architectures, training data, embodiment configurations, and benchmark-specific engineering. In this work, we introduce StarVLA-$α$, a simple yet strong baseline designed to study VLA design choices under controlled conditions. StarVLA-$α$ deliberately minimizes architectural and pipeline complexity to reduce experimental confounders and enable systematic analysis. Specifically, we re-evaluate several key design axes, including action modeling strategies, robot-specific pretraining, and interface engineering. Across unified multi-benchmark training on LIBERO, SimplerEnv, RoboTwin, and RoboCasa, the same simple baseline remains highly competitive, indicating that a strong VLM backbone combined with minimal design is already sufficient to achieve strong performance without relying on additional architectural complexity or engineering tricks. Notably, our single generalist model outperforms $π_{0.5}$ by 20\% on the public real-world RoboChallenge benchmark. We expect StarVLA-$α$ to serve as a solid starting point for future research in the VLA regime. Code will be released at https://github.com/starVLA/starVLA.

1 Introduction

VLA research aims to build general-purpose robotic agents, but fragmented architectures, data, embodiments, and engineering make performance comparisons difficult to interpret.

VLA models aim to move robotic manipulation beyond task-specific policies toward general-purpose robotic agents.
The field lacks clear evidence about which components drive performance gains because systems vary across architectures, pretraining data, embodiments, and benchmark-specific fine-tuning.
These differences entangle reported improvements with dataset choices, preprocessing pipelines, and benchmark-specific engineering.

StarVLA-𝛼

StarVLA-α is a controlled, simple baseline for studying VLA design choices under reduced experimental confounding. It tests whether streamlined modeling and unified evaluation can remain competitive across benchmarks and embodiments.

StarVLA-α reduces experimental confounding by using a pretrained VLM backbone without robot-specific pretraining or sophisticated action engineering.
The framework evaluates action head design, robot-specific pretraining, and data/interface engineering while keeping backbone, data scale, and training protocol identical.
A unified model is jointly trained across LIBERO, SimplerEnv, RoboTwin 2.0, and RoboCasa-GR1 without benchmark-specific adaptation.
StarVLA-α reaches leading performance on four benchmarks spanning five embodiments with a streamlined VLM design.
The analysis finds that added architectural and data-engineering complexity yields smaller and more context-dependent gains than often assumed.
A single generalist model can generalize across tasks and embodiments when trained jointly across benchmarks with strong initialization and a standardized pipeline.

2 StarVLA-α

StarVLA-α combines a unified Qwen3-VL backbone, minimal preprocessing, and a lightweight MLP action head in a common benchmark pipeline. It performs strongly across individual and unified multi-benchmark evaluations.

StarVLA-α tests whether a strong VLM backbone can deliver competitive performance without elaborate architectures or heavy data engineering.
The framework uses minimal data processing, including raw RGB images, language instructions, training-split action normalization, and official benchmark evaluation protocols.
Qwen3-VL processes vision and language inputs, while a lightweight MLP head regresses continuous action chunks from a designated action-token hidden state.
The unified pipeline integrates LIBERO, SimplerEnv, RoboTwin 2.0, and RoboCasa-GR1 while preserving consistent action representation and original benchmark protocols.
98.8% average success on LIBERO exceeds all previous approaches, while SimplerEnv improves by +6.8% on Google VM.
The unified generalist setup achieves competitive per-benchmark performance and notably improves results on challenging benchmarks such as RoboCasa-GR1.

3 Rethinking Common Practices in VLA Systems

StarVLA-α systematically tests action heads, action-specific pretraining, and data/interface engineering under a controlled setup. The results suggest that continuous action prediction and a strong VLM backbone matter more than additional architectural or pipeline complexity.

3.1 Do Different Action Head Designs Matter?: Continuous action prediction consistently outperforms discrete FAST-style prediction, while the three continuous action heads achieve comparable performance across benchmarks.All variants exceed 98% success on LIBERO and reach around 65% on WidowX.
3.1 Do Different Action Head Designs Matter?: Given a powerful VLM, a lightweight MLP action head is a simple, efficient, and competitive default.StarVLA-α achieves 53.8% on RoboCasa-GR1 despite using the simplest action-head design.
3.2 Does existing action-specific pretraining matter?: Additional action-specific pretraining does not consistently improve performance and can degrade results when the data is out of domain.The no-additional-pretraining baseline reaches 88.2 on RoboTwin 2.0 and 53.8 on RoboCasa-GR1 with sufficient task-specific data.
3.2 Does existing action-specific pretraining matter?: Pretraining can improve low-data performance when it closely matches the target task, but gains may not transfer across embodiments or tasks.InternData-A1 or RoboTwin pretraining improves RoboTwin performance yet reduces performance on RoboCasa.
3.3 Data and Interface Engineering: Data engineering techniques provide modest gains with limited task-specific data but little additional benefit once sufficient data is available.The evaluated choices include proprioception, history frames, delta actions, and relative actions, with other hyperparameters held fixed.

4 All-in-one Evaluation as a Generalist

The paper evaluates a single jointly trained generalist across multiple benchmarks and finds that a minimalist design remains competitive. Controlled analyses show that simple padding, moderate model scale, and larger batch sizes can support strong generalist performance.

Unified Evaluation: A single model jointly trained across LIBERO, SimplerEnv, RoboTwin 2.0, and RoboCasa-GR1 is evaluated without benchmark-specific fine-tuning.The generalist uses the combined training sets of all four benchmarks.
Unified Evaluation: The generalist model achieves state-of-the-art or competitive performance across most benchmarks and improves RoboCasa-GR1 performance by 3.5%.RoboCasa-GR1 contains 24 subtasks.
Action Parameterization: Simple padding matches RDT Action and the Multi-Action Head on LIBERO and RoboTwin 2.0 while improving Google Robot VM by 2.9% and RoboCasa-GR1 by 4.8%.All action vectors are uniformly expanded to 32 dimensions in the implementation setting.
Model Size: A 4B Qwen3-VL model is sufficient under the tested training scale, while the 8B model adds no substantial improvement over 4B, with gains within 1%.The 4B model also yields gains of 18.1% on WidowX and 6.6% on RoboCasa-GR1 relative to 2B.
Batch Size: Performance improves consistently with batch size, with batch size 512 reaching 57.3 on RoboCasa-GR1 and 57.2 on RoboTwin-Clean.The study evaluates batch sizes from 64 through 1024 under the same total training scale.

5 Real-World Experiments

The paper tests StarVLA-α on the standardized real-world RoboChallenge benchmark using physical robot experiments. On the ARX5 robot, StarVLA-α substantially exceeds π0.5 on both success rate and progress score.

Benchmark: RoboChallenge evaluates learned control policies on physical hardware using standardized tabletop manipulation tasks and repeated trials.Performance is measured by average success rate under predefined task-success criteria.
Results: On ARX5’s 11-task suite, StarVLA-α achieves a success rate of 33.6 and progress score of 54.5, versus π0.5’s 12.7 and 27.6.The comparison reports both success rate and progress score.

6 Related Works

Related work shows that VLA research spans diverse architectures, action-generation mechanisms, datasets, and preprocessing strategies. This diversity reflects rapid progress but also contributes to fragmentation across the field.

Vision-Language-Action Models: VLA policies build on large vision-language models and include decoupled vision-encoder–LLM pipelines, native multimodal models, and specialized action decoders.The cited approaches represent multiple architectural directions within the VLA paradigm.
Data and Action Parameterization: Robotic datasets require preprocessing to reconcile differences in control frequencies, camera viewpoints, and action formats.Related work also explores discretized token prediction, continuous autoregressive control, and action chunking.

7 Conclusion

The paper presents StarVLA-α as a simple VLA baseline that combines a strong VLM backbone with a lightweight action head and minimal data processing. Its results support using this simplified, reproducible framework as a starting point for future VLA research.

Conclusion: StarVLA-α combines a strong VLM backbone with a lightweight MLP action head and minimal data processing while achieving strong performance across benchmarks and real-world tasks.The design is intended to reduce architectural complexity and data-engineering demands.
Conclusion: Controlled experiments suggest that sophisticated action-head designs, heavy data engineering, and task-specific pretraining are not strictly necessary for generalist robot development.The simplified design is presented as reproducible and generalizable within the paper’s evaluated scope.

Vision-Language-Action Systems”

VLA research spans diverse action-modeling and data-engineering choices, from tokenized control to continuous generation and cross-embodiment preprocessing. These approaches support multimodal robotic control but leave a broad design space.

VLA systems map multimodal observations to deployable control signals using end-to-end frameworks.
Action modeling: Action modeling ranges from discretizing continuous controls into language tokens to continuous autoregression, action chunking, diffusion, and flow matching.
Data engineering: Robotic data pipelines also vary in normalization, proprioceptive-state injection, and padding for cross-embodiment alignment.

B Benchmark Details

The evaluation spans complementary simulation and real-world benchmarks covering multiple embodiments, task structures, and protocols. Training distinguishes single-embodiment specialists from multi-embodiment generalists under a largely benchmark-agnostic setup.

Benchmark coverage: The benchmark suite covers compositional learning, simulated real-world-policy evaluation, dual-arm coordination, humanoid tabletop manipulation, and standardized real-robot testing.
LIBERO: LIBERO contains 130 language-conditioned manipulation tasks across Spatial, Object, Goal, and Long suites, with LIBERO-Plus testing robustness under controlled shifts.
SimplerEnv: SimplerEnv evaluates real-world manipulation policies in simulation through standardized environments whose results correlate with physical robot performance.
RoboTwin 2.0 and RoboCasa-GR1: RoboTwin 2.0 tests bimanual coordination across 50 tasks, while RoboCasa-GR1 evaluates humanoid-style tabletop manipulation across 24 household-interaction tasks.
RoboChallenge: RoboChallenge evaluates policies directly on real robots under sensing noise, actuation uncertainty, and physical-environment variability.
Training setup: Default training initializes Qwen3-VL-4B, randomly initializes action heads, and trains directly on target benchmark data without action-specific pretraining.
Training paradigms: Specialists use one embodiment, whereas generalists jointly train on merged data from multiple embodiments and benchmark suites.

D.1 Effect of Model Initialization

Generalist VLA performance depends strongly on pretrained backbone quality, moderate model capacity, and large batches that expose diverse supervision during each optimization step. Scaling beyond moderate capacity yields diminishing returns, while batch size has a broad and consistent effect.

Model initialization: Pretrained VLM initialization consistently improves generalist performance, with stronger backbones generally producing better results across benchmarks.
Model initialization: Strong VLM initialization is a key ingredient for generalist training because better multimodal representations translate into stronger cross-benchmark generalization.
Model capacity: Increasing model size from 2B to 4B brings clear gains, whereas improvement from 4B to 8B is smaller and less consistent.
Model capacity: A 4B backbone captures most achievable gains and offers a favorable capacity-efficiency trade-off in this setup.
Batch size: Larger batch sizes steadily improve performance, with especially pronounced gains on SimplerEnv, RoboTwin 2.0, and RoboCasa-GR1.
Batch size: Batch size is a broader and more consistent optimization factor than model scaling for generalist VLA training.
Real-world benchmark: StarVLA-α consistently outperforms π0.5 and π0 on most platforms and task groups in both success rate and progress score.

F Real-world OOD Experiments

Real-world experiments test StarVLA-α under distribution shifts involving novel objects, colors, positions, and spatial coordinates. The model remains robust across three tasks, with OOD performance largely comparable to IID performance despite its simple unified framework.

Experimental setup: OOD experiments use a table-mounted Franka Research 3 with fixed third-person and wrist-mounted first-person RGB cameras.
Experimental setup: The three tasks test novel-object categorization, unseen colors and positions, and unseen row-column combinations in a 4 × 4 egg carton.
Results: StarVLA-α remains robust across all three tasks, with OOD performance largely comparable to IID performance without elaborate data engineering or task-specific tricks.
Results: Novel-object waste sorting reaches 85.0% OOD versus 87.5% in-domain success.
Results: The model achieves 68.0% success on unseen egg colors and 81.9% on unseen positions.
Evaluation settings: Simulation visualizations span WidowX, RoboCasa-GR1, Google Robot, and hard RoboTwin 2.0 scenes, while RoboChallenge and Franka deployments illustrate physical evaluation settings.
LIBERO-Plus: LIBERO-trained StarVLA-α variants transfer robustly to LIBERO-Plus perturbations and outperform prior baselines without robustness-specific augmentation.