VLANeXt: Recipes for Building Strong VLA Models

The paper introduces VLANeXt, a unified framework and practical recipe derived from systematically analyzing 12 key design choices across foundational components, perception, and action modeling, which enables the construction of state-of-the-art Vision-Language-Action models that outperform existing methods on both benchmarks and real-world tasks.

The Gist

Imagine you are trying to teach a robot to make a sandwich. In the past, you had to write a specific, rigid computer program for every single step: "move arm left 5cm," "grab bread," "move arm up." If the bread was slightly crooked, the robot would crash.

Then, researchers discovered VLAs (Vision-Language-Action models). Think of these as robots that have "read" the entire internet. They can look at a picture of a sandwich, understand your voice command ("Make me a sandwich"), and figure out the steps themselves. They are like a super-smart intern who has seen millions of videos of people making sandwiches.

However, building these robots has been chaotic. It's like a kitchen where every chef is using a different recipe, different measuring cups, and different ovens. Some chefs say, "Use more salt!" while others say, "No, less salt!" Because everyone is testing their recipes differently, no one knows which ingredients actually make the sandwich taste good.

Enter "VLANeXt": The Master Recipe Book

The authors of this paper decided to clean up the kitchen. They didn't just invent a new robot; they created a unified framework to test every possible ingredient and cooking method under the exact same conditions. They wanted to find the "secret sauce" that makes a VLA robot truly strong.

Here is the "recipe" they distilled, explained with simple analogies:

1. The Brain (The Foundation)

• The Old Way: The robot's brain was a bit lazy. It tried to use the same words it uses for talking to also figure out how to move its arm. It was like trying to write a poem and solve a math problem at the exact same time using the same scratchpad.
• The VLANeXt Fix: They gave the robot a dedicated "Action Brain" (a separate policy head). Think of it like hiring a specific "Movement Manager" who only focuses on moving the arms, while the "Language Manager" focuses on understanding your voice. They talk to each other, but they have their own jobs. This makes the robot much faster and more accurate.

2. The Eyes (Perception)

• The Old Way: The robot only looked at the world through one eye (a camera on the ceiling). If the lighting changed or an object was hidden, the robot got confused.
• The VLANeXt Fix: They gave the robot multiple eyes (a ceiling camera and a wrist camera). It's like wearing 3D glasses; the robot can now see depth and angles much better.
• The "Body Sense" Trick: The robot also learned to pay attention to its own body (proprioception). Instead of just looking at the sandwich, it "feels" where its arm is. The paper found that feeding this "body feeling" into the robot's main brain (the VLM) works better than just feeding it to the movement manager. It's like telling the chef, "I feel my hand is heavy," rather than just "Move your hand."

3. The Movement (Action Modeling)

• The Old Way: The robot tried to guess the next move one step at a time, like a person taking a single step and then stopping to think.
• The VLANeXt Fix:
  • Chunking: Instead of one step, the robot plans a "chunk" of 8 steps at once. It's like planning a whole sentence before speaking, rather than stuttering word-by-word.
  • Frequency Domain: This is the coolest part. The authors treated the robot's movements like a song. Just as a song has a rhythm and a melody, robot movements have patterns. By analyzing the "frequency" (the rhythm) of the movement, the robot can predict the future moves much more smoothly, like a DJ mixing tracks perfectly.

The Result: A Super-Intelligent Robot Chef

By combining these "recipes," the authors built VLANeXt.

• It's Smaller but Stronger: Even though VLANeXt is smaller (2.5 billion parameters) than some giants (7 billion parameters), it outperforms them. It's like a compact sports car that is faster than a massive truck because it's built with better engineering, not just more metal.
• It Handles Chaos: When the researchers tested it in a "LIBERO-plus" environment (where they randomly changed the lights, the background, or the robot's starting position), VLANeXt didn't panic. It adapted instantly, proving it truly understands the task, not just memorized the steps.
• Real-World Success: They tested it on real robots doing real tasks (cleaning tables, opening drawers, lifting baskets with two arms). It worked better than any previous model.

Why This Matters

Before this paper, building a robot was like trying to bake a cake without a recipe, guessing if you needed sugar or salt. VLANeXt provides the recipe.

The authors are giving away their "kitchen" (codebase) to everyone. Now, instead of every scientist reinventing the wheel, they can all use this solid foundation to build even better robots. They proved that you don't need a bigger, more expensive brain to be smart; you just need to connect the right parts in the right way.

In short: They took a chaotic, experimental field and turned it into a structured science, showing us exactly how to build robots that can actually help us in our daily lives.

Technical Summary

1. Problem Statement

The field of Vision-Language-Action (VLA) models has seen rapid growth, with numerous models proposed to enable general-purpose robot control by leveraging large foundation models (VLMs/LLMs). However, the current landscape is described as a "primordial soup":

• Fragmentation: Different research groups propose models with inconsistent training protocols, evaluation settings, and architectural choices.
• Lack of Clarity: It is difficult to determine which specific design choices (e.g., how to connect the VLM to the policy, how to model actions, or what inputs to use) truly drive performance.
• Inefficiency: Many approaches rely on aggressive model scaling or task-specific engineering rather than principled architectural design.

The paper aims to bring structure to this space by systematically re-examining the VLA design space under a unified framework and evaluation protocol to distill a practical "recipe" for building strong VLA models.

2. Methodology

The authors start with a simple baseline VLA (similar to RT-2 and OpenVLA) using an LLaMA backbone and SigLIP vision encoder. They systematically evolve this baseline along three dimensions, conducting ablation studies on the LIBERO and LIBERO-plus benchmarks.

A. Foundational Components

• Policy Module Design: Instead of reusing text tokens for action classification, the authors found that adding a separate policy head (decoupling action prediction from the linguistic space) yields better results. Furthermore, expanding the policy module from a single token to multiple tokens (MetaQuery-style) with deeper layers significantly improves performance.
• Action Chunking: Predicting a sequence of future actions (chunks) rather than single steps improves inference efficiency and action coherence. A chunk size of 8 was found optimal.
• Action Learning Objective: While the baseline used discrete binning (classification), the study found that continuous objectives perform better. Specifically, Flow Matching was adopted as it handles complex/multimodal distributions well, outperforming simple regression and diffusion-based losses in efficiency/accuracy trade-offs.
• VLM Backbone: Stronger VLM backbones correlate with better VLA performance. The authors selected Qwen3-VL-2B as a strong yet efficient choice, outperforming LLaMA and PaliGemma.
• VLM-Policy Connection: The study compared "loose" (decoupled), "tight" (layer-by-layer), and "soft" connections. The soft connection (layer-by-layer with learnable query buffers acting as a latent buffer) performed best, facilitating better representation transfer.

B. Perception Essentials

• Temporal History: Surprisingly, adding temporal observation history (past frames) degraded performance, likely introducing noise. The model performs best with the current frame only.
• Multi-View Inputs: Combining third-person and wrist (in-hand) camera views significantly improved performance by resolving spatial ambiguities.
• Proprioception Conditioning: The study compared injecting proprioception (robot state) into the VLM, the policy module, or both. Conditioning proprioception directly into the VLM yielded the best results, suggesting that fusing state information early with visual/language inputs is more effective than injecting it late into the policy.

C. Action Modeling Perspectives

• World Modeling: Adding an auxiliary objective to predict future image tokens improved performance but tripled training time, making it impractical for the final recipe.
• Time Series Forecasting (Frequency Domain): Inspired by time-series literature, the authors added a lightweight auxiliary loss that minimizes the Mean Squared Error (MSE) between predicted and ground-truth actions in the frequency domain (using Discrete Cosine Transform). This provided a significant performance boost with negligible computational overhead.

3. Key Contributions

1. Systematic Design Space Exploration: The paper provides the first comprehensive, unified ablation study of VLA design choices, moving beyond ad-hoc comparisons.
2. 12 Key Findings (The Recipe): The authors distilled 12 specific design principles, including:
   • Use a dedicated, deeper policy module with a soft connection to the VLM.
   • Employ action chunking and continuous flow-matching losses.
   • Condition proprioception in the VLM, not just the policy.
   • Use multi-view inputs but avoid redundant temporal history.
   • Apply frequency-domain regularization for action prediction.
3. VLANeXt Model: A new, simple, yet state-of-the-art VLA model derived from these principles. It uses a 2.5B parameter model (Qwen3-VL-2B + policy head), which is smaller than many competitors (e.g., OpenVLA-OFT at 7B) but achieves superior results.
4. Open Source: The authors release a unified, lightweight, and easy-to-use codebase to standardize VLA training and evaluation for the community.

4. Results

The VLANeXt model was evaluated on standard and robustness benchmarks:

• LIBERO Benchmark (Standard Tasks): VLANeXt achieved a 97.4% average success rate across Spatial, Object, Goal, and Long suites, outperforming previous SOTA methods like OpenVLA-OFT (97.1%) and π0 (86.0%).
• LIBERO-plus Benchmark (Robustness/Generalization): This benchmark introduces perturbations in lighting, background, robot state, and language. VLANeXt achieved an average success rate of 80.1%, significantly outperforming OpenVLA-OFT (69.6%) by over 10 percentage points. It demonstrated strong robustness across all perturbation types.
• Real-World Experiments: Tested on a Franka Emika arm (single-arm) and Aloha system (bimanual) on tasks like table cleaning and drawer manipulation. VLANeXt achieved the highest success rates (e.g., 14/20 on single-arm cleaning vs. 7/20 for OpenVLA-OFT), demonstrating effective cross-embodiment adaptability even without specific bimanual pre-training.

5. Significance

• Paradigm Shift: The paper argues that strong VLA performance does not necessarily require massive model scaling or complex world models. Instead, principled design choices within a unified framework are the key drivers of performance.
• Efficiency: VLANeXt achieves SOTA performance with a smaller model size (2.5B vs. 7B+), proving that architectural efficiency matters.
• Community Standard: By releasing a unified codebase and a clear "recipe," the authors aim to move the field away from fragmented, hard-to-reproduce experiments toward systematic, controlled exploration of the VLA design space.
• Insightful Findings: The discovery that proprioception should be conditioned in the VLM and that temporal history can be detrimental challenges common assumptions in the field, offering new directions for future research.