Less is More: Reducing Complexity in Vision-Language-Action Systems

Conference: ECCV 2026
Paper: ECCV Official
Code: https://github.com/starVLA/starVLA
Area: Robotics
Keywords: Embodied AI, Vision-Language-Action Models, Qwen3-VL, Cross-Embodiment Generalization, Minimalist System Design

TL;DR

Targeting the fragmentation, benchmark-specific overfitting, and bloated engineering in Vision-Language-Action (VLA) systems, this paper introduces StarVLA-α, demonstrating that a strong pretrained VLM backbone paired with a lightweight MLP continuous action head and simple zero-padding matches or outperforms state-of-the-art complex architectures across multiple simulation benchmarks and real-world physical robot evaluations.

Background & Motivation

Vision-Language-Action (VLA) models have emerged as the foundational paradigm for building general-purpose robotic manipulation policies. Driven by foundational vision-language models (VLMs), web-scale multimodal supervision, and growing robot demonstration datasets, recent systems have demonstrated impressive zero-shot task transfer and spatial execution capabilities. However, unlike pure vision-language modeling where recipes have converged toward standardized pretraining and alignment protocols, the VLA research landscape has grown increasingly fragmented and excessively complex. Existing models diverge across visual encoder selections, action decoding paradigms (such as autoregressive tokenization, diffusion models, and flow-matching policies), dual-system hierarchical modules, proprioceptive state injection, history frame stacking, and task-specific coordinate transformations.

This fragmentation creates a severe attribution dilemma: empirical gains reported across various papers are heavily entangled with proprietary dataset selections, ad-hoc preprocessing tricks, and benchmark-specific hyperparameter tuning. Consequently, it remains obscure whether performance improvements stem from genuine modeling breakthroughs or from benchmark-tailored engineering and evaluation variance. Furthermore, complex modules meticulously engineered for specific platforms frequently suffer catastrophic performance drops when confronted with out-of-distribution environmental shifts or new robot embodiments, imposing high reproducibility barriers and hindering methodological consensus across the robotics community.

To cut through this systemic complexity, this work adopts an Occam's razor perspective to determine what actually drives VLA capabilities under strictly controlled conditions. Core idea: introduce StarVLA-α, an intentionally minimalist yet robust baseline that pairs a pretrained Qwen3-VL backbone with a lightweight MLP continuous action head, systematically demonstrating that specialized action heads, large-scale robotic pretraining, and heavy data engineering provide limited, context-dependent gains, while a simple zero-padding strategy enables effective cross-embodiment multi-task generalist policies.

Method

Overall Architecture

StarVLA-α is grounded in the "minimal-sufficiency hypothesis": a sufficiently strong VLM backbone combined with a simple continuous action head captures virtually all the performance gains previously attributed to complex architectural designs. The framework eliminates specialized visual feature encoders, multimodal fusion bridges, and separate action expert networks. At the input stage, raw single-frame RGB images and natural language instructions are directly processed by the unified Qwen3-VL backbone without custom visual cropping, temporal history stacking, or proprioceptive state concatenation. At the output stage, hidden states from a designated action token are extracted and fed directly into a lightweight Multi-Layer Perceptron (MLP) head to regress continuous action chunks in a single forward pass. For cross-embodiment generalist training, action dimensions across disparate robots are aligned via naive zero-padding, allowing a single unified model to be trained jointly without embodiment-specific heads.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Multi-Embodiment Multi-Task Input<br/>Raw single-frame RGB + language task instruction"] --> B["Zero-Redundancy Data Flow & Minimal Action Normalization<br/>No proprioception / history frames; zero-mean unit-variance on train split"]
    B --> C["Unified VLM Backbone & Continuous Action Head<br/>Qwen3-VL multimodal hidden states + lightweight MLP regressing action chunk"]
    C --> D["Cross-Embodiment Unified Zero-Padding Representation<br/>Pad robot action vectors to 32D; joint training across all datasets"]
    D --> E["Multi-Embodiment Control Execution<br/>Deploy directly to single-arm, dual-arm, wheeled, and humanoid platforms"]

Key Designs

1. Zero-Redundancy Data Flow & Minimal Action Normalization: Stripping extraneous inputs and complex coordinate schemes

Conventional VLA practices widely assume that robotic policies must rely on proprioceptive joint states and stacked temporal history frames, while expending substantial effort engineering coordinate spaces between absolute, delta, and relative end-effector representations. Such engineering adds platform-dependent incompatibilities and introduces temporal alignment vulnerabilities. StarVLA-α completely strips these auxiliary inputs: it neither concatenates history frames nor ingests proprioceptive state vectors, relying purely on the current RGB observation and task language instruction. Actions are simply normalized to zero mean and unit variance using the training split statistics, adhering strictly to official benchmark protocols during inference. Empirical ablations confirm that modern VLMs possess sufficient spatial and semantic comprehension to infer operational dynamics directly from appearance, and as demonstration scale reaches reasonable thresholds, the marginal gains of manual data engineering vanish entirely.

2. Unified VLM Backbone & Continuous Action Head: Replacing specialized action experts with a strong VLM and a lightweight MLP

Prior literature features heavily contested action decoding paradigms, spanning autoregressive discrete token prediction (e.g., FAST style), diffusion and flow-matching continuous generation (e.g., \(\pi_0\) style), and decoupled dual-system architectures separating high-level reasoning from low-level execution (e.g., GR00T style). These alternatives impose steep training overheads and inference latency. StarVLA-α adopts an open-source Qwen3-VL-4B model as a native multimodal foundation, inherently unifying visual and textual token representations without alignment loss across mismatched encoders. A dedicated Action Token is appended to the input sequence; the corresponding output hidden state \(\mathbf{h}_{\text{act}}\) is extracted from the backbone and projected through a lightweight MLP to regress continuous action chunks \(\hat{\mathbf{A}} = \text{MLP}(\mathbf{h}_{\text{act}})\). This eliminates iterative denoising, multi-step sampling, and auxiliary cross-attention modules, demonstrating that continuous action regression directly from a capable VLM is both computationally optimal and highly competitive.

3. Cross-Embodiment Unified Zero-Padding Representation: Replacing multi-action heads with naive dimensional padding

When developing a unified generalist agent across diverse embodiments—ranging from single-arm manipulators (Franka, UR5) to dual-arm wheeled platforms and full-scale humanoid robots (Fourier GR1)—traditional architectures (such as ABot-VLA, LingBot-VLA, or RDT) engineer complex multi-action heads or elaborate shared latent manifold mappings. These designs fragment parameter capacity and scale poorly when onboarding new hardware. StarVLA-α discards embodiment-specific head routing in favor of an elementary zero-padding mechanism: it designates a maximum action dimension \(D_{\max} = 32\) that accommodates all supported platforms, padding unconstrained degrees of freedom with zeros for lower-DoF robots. Leveraging the high parameter capacity and visual reasoning of Qwen3-VL, the model intrinsically resolves embodiment identity and kinematic constraints from visual cues and task context, enabling a single generalist model to optimize jointly across heterogeneous robot data without structural modifications.

Loss & Training

StarVLA-α is trained end-to-end via standard supervised imitation learning. Given ground-truth action chunks \(\mathbf{A} \in \mathbb{R}^{T \times D_{\max}}\) and predicted chunks \(\hat{\mathbf{A}} = \text{MLP}(\mathbf{h}_{\text{act}})\), the network is optimized using mean squared error (MSE) loss across all valid control dimensions:

\[\mathcal{L}_{\text{action}} = \frac{1}{T \cdot D_{\text{valid}}} \sum_{t=1}^{T} \sum_{d=1}^{D_{\text{valid}}} \|\mathbf{A}_{t, d} - \hat{\mathbf{A}}_{t, d}\|^2\]

In the all-in-one Generalist training regime, the unified model is trained across five pooled datasets spanning LIBERO, SimplerEnv, RoboTwin 2.0, and RoboCasa-GR1 using a global batch size of 512 and a learning rate of \(1 \times 10^{-4}\) without per-task tuning. Batch size scaling analyses confirm that large batch sizes (512 to 1024) provide crucial batch-level sample diversity that prevents catastrophic gradient conflict across heterogeneous robotic embodiments.

Key Experimental Results

Main Results

StarVLA-α is evaluated against established SOTA VLA baselines on four major benchmarks: LIBERO (lifelong knowledge transfer), SimplerEnv (real-to-sim transfer across WidowX and Google Robot), RoboTwin 2.0 (dual-arm coordinated manipulation), and RoboCasa-GR1 (humanoid mobile manipulation across 24 kitchen tasks).

Paradigm	Method	LIBERO avg (%)	Simpler WidowX (%)	Simpler Google VM (%)	RoboTwin 2.0 Clean* (%)	RoboCasa-GR1 (24-task avg, %)
Specialist	OpenVLA-OFT	97.1	31.3	63.0	–	–
Specialist	\(\pi_0\)	94.1	27.1	58.8	65.9	–
Specialist	\(\pi_0\)+FAST	85.5	39.5	61.9	–	–
Specialist	\(\pi_{0.5}\)	96.9	46.9	72.7	82.7	37.0
Specialist	GR00T-N1.6	97.0	62.0	67.7	–	47.6
Specialist	StarVLA-α (Ours)	98.8	64.6	76.0	88.2	53.8
Generalist	StarVLA-α Generalist (Ours)	97.8	65.2	74.3	88.7	57.3

Note: Clean* denotes training with both clean and randomized trajectories on RoboTwin 2.0; official evaluation protocols are strictly followed across all benchmarks.

Ablation Study

Systematic controlled ablations isolate the impact of action head architectures, action pretraining datasets, data engineering choices, and embodiment parameterizations.

Table 1: Controlled comparison of action head designs under identical Qwen3-VL backbone and data settings

Action Head Variant	Core Decoding Mechanism	LIBERO avg (%)	Simpler WidowX (%)	Simpler Google VM (%)	RoboCasa-GR1 (24-task avg, %)
StarVLA-α-FAST	Autoregressive discrete tokenization	97.8	35.6	60.1	45.0
StarVLA-α-\(\pi\)	Flow-matching continuous generation	98.0	65.9	76.6	48.9
StarVLA-α-GR00T	Dual-system VLM + flow matching	98.6	65.3	75.3	52.8
StarVLA-α (MLP)	Direct continuous regression via MLP	98.8	64.6	76.0	53.8

Table 2: Comparison of multi-embodiment action parameterizations in the Generalist training setting

Parameterization Scheme	Design Description	LIBERO Avg (%)	Simpler WidowX (%)	Simpler Google VM (%)	RoboTwin 2.0 Random* (%)	RoboCasa-GR1 Avg (%)
RDT Action	Continuous unified manifold alignment	97.2	63.9	71.4	86.6	52.3
Multi-Action Head	Dedicated decoupled head per embodiment	97.2	60.6	67.8	86.1	53.5
Simple Padding (Ours)	Unified zero-padding to 32D with single head	97.8	65.2	74.3	87.8	57.3

Table 3: Real-world physical robot evaluation on the standardized RoboChallenge benchmark (ARX5 robot, 11 tasks)

Model	Success Rate (SR, %)	Progress Score (Score)	Highlight Tasks (Arrange Flowers / Place Shoes / Cup on Coaster)
\(\pi_0\)	3.6	14.7	0.0% / 0.0% / 0.0%
\(\pi_{0.5}\)	12.7	27.6	0.0% / 0.0% / 70.0%
StarVLA-α (Ours)	33.6	54.5	40.0% / 50.0% / 100.0%

Key Findings

• Continuous action prediction is critical, but complex continuous heads yield minimal marginal returns: Switching from discrete autoregressive tokenization (FAST) to continuous action prediction produces major performance leaps (e.g., from 35.6% to 64.6% on WidowX, and from 45.0% to 53.8% on RoboCasa). However, among continuous decoders, the lightweight MLP head performs on par with or outperforms sophisticated flow-matching models (\(\pi_0\)) and dual-system modules (GR00T), proving that complex generative heads are not inherently superior when conditioned on a capable VLM backbone.
• Heterogeneous robotic pretraining acts as a double-edged sword: Incorporating massive out-of-domain robot datasets such as Open X-Embodiment (OXE) before downstream fine-tuning degrades performance significantly on RoboTwin (plunging from 50.3% to 30.2%) and RoboCasa (from 9.8% to 1.2% in low-data regimes). Even in-domain robotic pretraining that aids a specific target embodiment impairs cross-embodiment generalization to distinct robots.
• Generalist training unleashes positive transfer in complex tasks: Rather than suffering from negative interference, training a single model jointly across all four benchmarks improves performance on the most demanding humanoid benchmark (RoboCasa-GR1), boosting the 24-task average success rate from 53.8% (Specialist) to 57.3% (Generalist).
• Model capacity exhibits a clear saturation threshold: Scaling the VLM backbone from 2B to 4B yields substantial gains (+18.1% on WidowX, +6.6% on RoboCasa-GR1), but scaling further from 4B to 8B provides negligible improvement (<1%), identifying 4B as the current sweet spot for compute efficiency and policy performance.

Highlights & Insights

• Minimalist Occam's razor cuts through VLA folklore: Demonstrates that modern pretrained VLMs inherently possess the spatial-semantic reasoning required for robotic manipulation; complex action generation modules, proprioceptive vectors, and temporal frame stacking are largely redundant when supported by a strong foundation model.
• Radical simplicity in cross-embodiment alignment: Shows that zero-padding action vectors to a maximum dimension of 32 allows the VLM to naturally disambiguate robot kinematics and control semantics from visual observations, obsoleting complex multi-action routing networks.
• Transparent and reproducible benchmarking: Establishes StarVLA as an open-source, modular codebase that eliminates hidden confounders, providing a standardized, controlled platform for future embodied AI investigations.

Limitations & Future Work

• Lack of explicit tactile and force feedback: The architecture operates entirely on passive RGB visual observations without dynamic tactile or closed-loop force sensing, limiting applicability to high-precision contact-rich industrial assembly or forceful insertion.
• Action trajectory smoothness under chunk boundary transitions: Direct regression of open-loop action chunks without temporal re-planning or overlapping diffusion blending may occasionally introduce acceleration discontinuities at chunk boundaries, which could be mitigated via lightweight temporal filtering.
• Physical hardware diversity: While evaluated extensively across 11 tasks on the ARX5 tabletop arm in RoboChallenge, real-world evaluations have not yet been extended to dynamic bipedal humanoid locomotion or high-speed mobile manipulation in unstructured outdoor settings.

Related Work & Insights

• vs OpenVLA / OpenVLA-OFT: OpenVLA relies on discretized action tokenization and heavy OXE pretraining, incurring high sequence lengths and discretization errors; OpenVLA-OFT transitions to continuous action tokens but retains heavy pretraining overhead. StarVLA-α shows that native Qwen3-VL features paired with a direct MLP head outperform OpenVLA variants without requiring any action-specific pretraining.
• vs \(\pi_0\) / \(\pi_{0.5}\): \(\pi\)-series policies advocate for continuous flow-matching action experts and massive web-and-robot co-training. StarVLA-α demonstrates that under identical backbones, a simple MLP action head achieves equivalent or superior success rates at a fraction of the computational and architectural complexity.
• vs GR00T-N1.5 / N1.6: GR00T couples a VLM (System 2) with a flow-matching action executor (System 1) and custom embodiment-specific interfaces. StarVLA-α proves that an integrated end-to-end VLM with elementary zero-padding achieves superior performance on humanoid tasks (RoboCasa-GR1) while dramatically reducing system complexity.

Rating

• Novelty: ⭐⭐⭐⭐☆ While deliberately avoiding architectural complexity, its rigorous methodology, controlled empirical deconstruction, and decisive falsification of prevalent design assumptions offer profound scientific value to the community.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comprehensive coverage across four major simulation benchmarks (LIBERO, SimplerEnv, RoboTwin 2.0, RoboCasa-GR1), real-world physical evaluations on RoboChallenge, and deep controlled ablations across model scales, batch sizes, and data engineering.
• Writing Quality: ⭐⭐⭐⭐⭐ Extremely lucid, logical narrative with clear motivation, insightful takeaways, and well-structured empirical evidence directly addressing critical bottlenecks in VLA research.
• Value: ⭐⭐⭐⭐⭐ Serves as a vital benchmark baseline and conceptual corrective for the rapidly proliferating embodied AI ecosystem, anchoring future research on solid, reproducible foundations.