StarVLA: A Lego-like Codebase for Vision-Language-Action Model Developing

StarVLA Community

cs.ROApr 6, 2026

StarVLA: A Lego-like Codebase for Vision-Language-Action Model Developing

Authors: StarVLA Community

Organizations: Von Neumann Institute, HKUST

Abstract

Building generalist embodied agents requires integrating perception, language understanding, and action, which are core capabilities addressed by Vision-Language-Action (VLA) approaches based on multimodal foundation models, including recent advances in vision-language models and world models. Despite rapid progress, VLA methods remain fragmented across incompatible architectures, codebases, and evaluation protocols, hindering principled comparison and reproducibility. We present StarVLA, an open-source codebase for VLA research. StarVLA addresses these challenges in three aspects. First, it provides a modular backbone--action-head architecture that supports both VLM backbones (e.g., Qwen-VL) and world-model backbones (e.g., Cosmos) alongside representative action-decoding paradigms, all under a shared abstraction in which backbone and action head can each be swapped independently. Second, it provides reusable training strategies, including cross-embodiment learning and multimodal co-training, that apply consistently across supported paradigms. Third, it integrates major benchmarks, including LIBERO, SimplerEnv, RoboTwin~2.0, RoboCasa-GR1, and BEHAVIOR-1K, through a unified evaluation interface that supports both simulation and real-robot deployment. StarVLA also ships simple, fully reproducible single-benchmark training recipes that, despite minimal data engineering, already match or surpass prior methods on multiple benchmarks with both VLM and world-model backbones. To our best knowledge, StarVLA is one of the most comprehensive open-source VLA frameworks available, and we expect it to lower the barrier for reproducing existing methods and prototyping new ones. StarVLA is being actively maintained and expanded; we will update this report as the project evolves. The code and documentation are available at https://github.com/starVLA/starVLA.

Figures & tables

Figure 1: Conceptual view of the unified VLA formulation adopted in StarVLA. A policy $\pi$ maps visual observations and a language instruction to a future action chunk. The training objective decomposes as $\mathcal{L}=\mathcal{L}_{\mathrm{action}}+\mathcal{L}_{\mathrm{aux}}$ , where different model families correspond to different forms of $\mathcal{L}_{\mathrm{aux}}$ .

Figure 2: Overview of four representative approaches for adapting Vision-Language Models into Vision-Language-Action frameworks in StarVLA (FAST, OFT, $\pi$ , and GR00T) under a unified interface.

Figure 3: Overview of the StarVLA framework. We present a unified and modular pipeline that connects heterogeneous data sources, pluggable dataloaders, and flexible data representations with a standardized model forwarding interface. The framework supports diverse vision-language foundation models and VLA architectures, enabling end-to-end training and deployment.

Google Robot	Models	Pick Coke Can	Move Near	Open/Close Drawer	Open Top Drawer and Place Apple	Avg
Visual Matching	RT-1 Brohan et al. (2022)	85.7	44.2	73.0	6.5	52.4
	RT-1-X Collaboration et al. (2023)	56.7	31.7	59.7	21.3	42.4
	RT-2-X Brohan et al. (2023)	78.7	77.9	25.0	3.7	46.3
	OpenVLA Kim et al. (2024)	18.0	56.3	63.0	0.0	34.3
	CogACT Li et al. (2024a)	91.3	85.0	71.8	50.9	74.8
	SpatialVLA Qu et al. (2025)	86.0	77.9	57.4	-	75.1

Table 4: Detailed results on the SimplerEnv Google Robot benchmark. Numbers are officially reported unless marked with

*

, which denotes our reimplementation. We report StarVLA-OFT with Qwen3-VL-4B as a representative configuration due to the high evaluation cost on this platform.

Method	Avg (%)	Method	Avg (%)
$\pi_{0.5}$ Intelligence et al. (2025b)	37.0	GR00T-N1.6 Bjorck et al. (2025)	47.6
StarVLA-FAST	39.0	StarVLA- $\pi$	43.9
StarVLA-GR00T	47.8	StarVLA-OFT	48.8

Table 5: Average success rate on RoboCasa-GR1 (24 tasks) under the single-benchmark training setting.

Method	Clean	Random	Method	Clean	Random
$\pi_{0}$ Black et al. (2024)	65.9	58.4	$\pi_{0.5}$ Intelligence et al. (2025b)	82.7	76.8
X-VLA Zheng et al. (2025a)	72.9	72.8	Lingbot-VLA Wu et al. (2026)	88.6	86.7
StarVLA-FAST	72.5	83.2	StarVLA-OFT	88.2	88.3
StarVLA-GR00T	88.0	88.5	StarVLA- $\pi$	88.1	88.8

Table 7: Detailed results on the RoboTwin 2.0 benchmark. We report different StarVLA model architecture on this platform.

Figure 4: Perception–action co-optimization dynamics under different co-training strategies (reproduced from a StarVLA-based spatially guided co-training study ST4VLA ( Ye et al., 2026a ) ). From left to right: (a) spatial grounding performance ([email protected] on RefCOCO-g); (b) manipulation success rate (WidowX); (c) gradient subspace alignment (PSS) between spatial grounding and action objectives under vanilla co-training vs. spatially guided co-training.

Per-GPU batch	Global batch	Time / 100K steps	Seconds / step	Samples / s	GPU util
2	16	19:32:17	0.703	22.7	74%
4	32	24:35:59	0.886	36.1	89%
8	64	31:25:38	1.131	56.6	92%
16	128	49:15:53	1.774	72.2	91%
24	192	66:47:02	2.404	79.9	96%

Table 10: Single-node training efficiency (8

\times

A100). Sample throughput is derived from the measured time per 100K steps and the global batch size.

Figure 5: Per-step latency and throughput on a single 8-GPU node. Left: step latency as a function of per-GPU batch size for our method on A100 and H200, compared with LingBot-VLA and Dexbotic (both on 8×H200). Right: training throughput and GPU utilization on 8×A100 across batch sizes.

# GPUs	Global batch	Time / 100K steps	Seconds / step	Samples / s	Scaling eff.
8	64	20:25:48	0.735	87.0	100%
16	128	23:36:00	0.850	150.7	86.7%
32	256	24:58:45	0.899	284.7	81.9%
64	512	25:40:59	0.925	553.8	79.6%
128	1024	25:35:26	0.921	1111.5	79.9%
256	2048	25:51:41	0.931	2200.0	79.1%

Table 11: Multi-node training efficiency (per-GPU batch = 8). “Ideal” scaling assumes linear growth of samples/s from the 8-GPU baseline.

Figure 6: Multi-node scaling efficiency. Left: per-step latency rises noticeably from 8 to 32 GPUs due to inter-node communication overhead, then plateaus between 64 and 256 GPUs. Right: measured sample throughput versus ideal linear scaling; parallel efficiency stabilizes around 79–80% beyond 32 GPUs.

Explore similar work

Apr 13, 2026

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

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.

Jinhui Ye, Ning Gao, Senqiao Yang +7

HKUST · XJTU · CUHK +3

Robot Foundation Models Robotic Manipulation Vision-Language-Action Models Robot Learning

Ask PDF

Apr 21, 2026

VLA Foundry: A Unified Framework for Training Vision-Language-Action Models

We present VLA Foundry, an open-source framework that unifies LLM, VLM, and VLA training in a single codebase. Most open-source VLA efforts specialize on the action training stage, often stitching together incompatible pretraining pipelines. VLA Foundry instead provides a shared training stack with end-to-end control, from language pretraining to action-expert fine-tuning. VLA Foundry supports both from-scratch training and pretrained backbones from Hugging Face. To demonstrate the utility of our framework, we train and release two types of models: the first trained fully from scratch through our LLM-->VLM-->VLA pipeline and the second built on the pretrained Qwen3-VL backbone. We evaluate closed-loop policy performance of both models on LBM Eval, an open-data, open-source simulator. We also contribute usability improvements to the simulator and the STEP analysis tools for easier public use. In the nominal evaluation setting, our fully-open from-scratch model is on par with our prior closed-source work and substituting in the Qwen3-VL backbone leads to a strong multi-task table top manipulation policy outperforming our baseline by a wide margin. The VLA Foundry codebase is available at https://github.com/TRI-ML/vla_foundry and all multi-task model weights are released on https://huggingface.co/collections/TRI-ML/vla-foundry. Additional qualitative videos are available on the project website https://tri-ml.github.io/vla_foundry.

Jean Mercat, Sedrick Keh, Kushal Arora +5

Toyota Research Institute

Robotic Manipulation Vision-Language-Action Models VLMs for Robotics

Ask PDF

Feb 26, 2026

Rethinking the Practicality of Vision-language-action Model: A Comprehensive Benchmark and An Improved Baseline

Vision-Language-Action (VLA) models have emerged as a generalist robotic agent. However, existing VLAs are hindered by excessive parameter scales, prohibitive pre-training requirements, and limited applicability to diverse embodiments. To improve the practicality of VLAs, we propose a comprehensive benchmark and an improved baseline. First, we propose CEBench, a new benchmark spanning diverse embodiments in both simulation and the real world with consideration of domain randomization. We collect 14.4k simulated trajectories and 1.6k real-world expert-curated trajectories to support training on CEBench. Second, using CEBench as our testbed, we study three critical aspects of VLAs' practicality and offer several key findings. Informed by these findings, we introduce LLaVA-VLA, a lightweight yet powerful VLA designed for practical deployment on consumer-grade GPUs. Architecturally, it integrates a compact VLM backbone with multi-view perception, proprioceptive tokenization, and action chunking. To eliminate reliance on costly pre-training, LLaVA-VLA adopts a two-stage training paradigm including post-training and fine-tuning. Furthermore, LLaVA-VLA extends the action space to unify navigation and manipulation. Experiments across embodiments demonstrate the capabilities of generalization and versatility of LLaVA-VLA , while real-world mobile manipulation experiments establish it as the first end-to-end VLA model for mobile manipulation. We will open-source all datasets, codes, and checkpoints upon acceptance to foster reproducibility and future research.

Wenxuan Song, Jiayi Chen, Xiaoquan Sun +12

The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China. · Huazhong University of Science and Technology, Wuhan, China. · Westlake University, Hangzhou, China. +1

Vision-Language-Action Models Efficient VLA Models Mobile Manipulation

Ask PDF

Abstract

Explore similar work

StarVLA-ααα: Reducing Complexity in Vision-Language-Action Systems

VLA Foundry: A Unified Framework for Training Vision-Language-Action Models

Rethinking the Practicality of Vision-language-action Model: A Comprehensive Benchmark and An Improved Baseline

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