VLANeXt: Recipes for Building Strong VLA Models

• VLANeXt achieves higher success rates than prior methods on both LIBERO and LIBERO-plus despite its smaller model size.The comparison spans standard task performance and robustness/generalization across model scales.Source: Figure 1 · p. 1, §1 · p. 2
• The study conducts more than 500 experiments across foundational components, perception essentials, and action modeling, distilling 12 key findings.Variants are evaluated on LIBERO and LIBERO-plus, with LIBERO-plus adding controlled and unseen perturbations for robustness and generalization.Source: §1 · p. 2
• The design recipe finds that soft VLM-policy coupling slightly outperforms loose and tight coupling, while video inputs fail to provide useful action-learning information.It also finds that conditioning proprioception in the VLM is better than omitting it or injecting it directly into the policy module.Source: §1 · p. 2
• Treating action generation as time-series forecasting and adding frequency-domain modeling improves action prediction effectively and efficiently.Source: §1 · p. 2
• Beyond the core study, the work explores smaller VLA backbones, latent action pretraining, and World Action Models.A smaller VLANeXt variant remains competitive while improving efficiency, and latent action pretraining uses broader robot trajectories before supervised action fine-tuning.Source: §1 · p. 2
• The released codebase standardizes training and evaluation while enabling reproduction, shared-protocol comparisons, and development of new VLA variants.Source: §1 · p. 3

2. Recipes for Building Strong VLA Models

• The roadmap evaluates a minimal 3B-parameter LLaMA 3.2–SigLIP2 RT-2-style baseline on LIBERO and LIBERO-plus, primarily using the spatial suite.Actions are represented as discretized bin indices predicted through reused text tokens, providing a reference point for design comparisons.Source: §2 · p. 3

2.1. The Foundational Components

• Replacing one class token with 16 tokens and expanding the policy network from 2 to 12 layers improves performance; 8-action chunking, flow matching, and Qwen3-VL-2B are adopted.Stronger VLMs follow Qwen3-VL-4B > Qwen3-VL-2B > LLaMA-3.2-3B and PaliGemma-3B, with Qwen3-VL-2B chosen as a strong yet efficient model.Source: §2.1 · pp. 3–4

2.2. The Perception Essentials

• Multi-view third-person plus wrist observations and VLM-side proprioception improve performance, whereas adding temporal observation history slightly degrades action generation.A transformer projector performs slightly better than alternatives, but the final design uses a linear projector for simplicity.Source: §2.2 · pp. 4, 6

2.3. Action Modelling Perspectives

• A frequency-domain auxiliary loss improves action generation slightly beyond world modelling with negligible overhead, while world modelling improves performance at substantially higher computational cost.The frequency loss uses a discrete cosine transform and emphasizes low-frequency components to regularize trajectory jitter and improve generalization.Source: §2.3 · p. 6

2.4. Summary of Recipes

• VLANeXt combines deeper policy modelling, action chunking, flow matching, Qwen3-VL-2B, soft layer-wise VLM-policy connections, multi-view inputs, VLM-side proprioception, and frequency-domain regularization.World modelling also improves performance but nearly triples training time, so it is excluded from the final recipe.Source: §2.4 · pp. 6–7, §2.3 · p. 6

3.1. Settings

• Experiments fine-tune all models on LIBERO for 10,000 steps with batch size 256, then evaluate corresponding suites on both LIBERO and LIBERO-plus.Standard LIBERO contains Spatial, Object, Goal, and Long suites, while LIBERO-plus tests unseen perturbations without perturbation training.Source: §3.1 · p. 7

3.2. LIBERO Benchmark Results

• VLANeXt achieves state-of-the-art performance on LIBERO against direct policy-learning and pretrained-VLM-based VLA methods.The comparison includes Diffusion Policy, Octo, MDT, OpenVLA (Kim et al., 2024a), π0 (Black et al., 2024), and OpenVLA-OFT (Kim et al., 2025), among others.Source: §3.2 · p. 7

3.3. LIBERO-plus Benchmark Results

• VLANeXt demonstrates strong generalization to unseen perturbations, improving success rate by 13% over OpenVLA-OFT (Kim et al., 2025) on LIBERO-plus.LIBERO-plus introduces systematic evaluation variations involving visual, physical, and semantic perturbations.Source: §3.3 · p. 7, §3.1 · p. 7

4.1. Settings

• The evaluation covers four tasks: single-arm table cleaning and drawer manipulation, plus bimanual basket lifting and bimanual table cleaning.Single-arm experiments use Franka Emika, while bimanual experiments use the Aloha system.Source: §4.1 · pp. 7–8
• Training uses 50 episodes per task, 20 evaluation trials, 100k-step DROID pretraining, and 20k-step task fine-tuning at 1 × 10^-4.For bimanual adaptation, the proprioception projector and final action-generation layer are reinitialized while other pretrained weights remain fixed.Source: §4.1 · p. 8

4.2. Results

• Compared with OpenVLA-OFT (Kim et al., 2025) and π0 (Black et al., 2024) under matched fine-tuning, VLANeXt performs well in real-world experiments and adapts decently to bimanual tasks without bimanual training.The authors interpret these results as evidence that their recipes yield a strong, deployable VLA with cross-embodiment adaptability.Source: §4.2 · p. 8

Model architectures

• The codebase now supports smaller, larger, and video-generation backbones, including Qwen3.5-0.8B, Qwen3-VL-4B, Qwen3-VL-8B, and WAN2.2-5B.These options extend exploration from compact and large VLMs to World Action Model-style video-generation architectures.Source: §5 · p. 8

Action and auxiliary objectives

• VLANeXt adds x0-space action generation, FAST frequency-domain action-token classification, video-generation losses, DINO latent image prediction, and language-action pretraining.Video losses support fast and joint modes, depending on whether action learning can attend to future video tokens.Source: §5 · pp. 8–9

Training strategies

• Latent-action pretraining uses all four LIBERO suites before suite-specific fine-tuning, with VAE, VQ-VAE, and VICReg formulations for constraining latent actions.The other three suites provide additional unlabeled pretraining data during each suite’s fine-tuning.Source: §5 · p. 9

Data loading and evaluation

• The upgraded codebase supports TFDS and LeRobot data, sequential or random episode loading, parallel benchmark rollouts, and EMA checkpoint saving.These additions target broader data compatibility and faster, more robust evaluation.Source: §5 · p. 9

Empirical observations

• Selected experiments show that larger VLAs can improve performance, latent-action pretraining benefits from cross-suite unlabeled data, VQ-VAE constraints perform best, and image or video modeling can aid action learning.VLANeXt-XL does not surpass VLANeXt-L because LIBERO lacks sufficient data to fine-tune an 8B model; DINO representations may facilitate visual-to-action knowledge transfer.Source: §5 · pp. 9–10, Table 5 · p. 10

Released baselines

• The release includes VLANeXt-LAM, VLANeXt-S, VLANeXt-X, VLANeXt-XL, VLANeXt-JEPA, and VLANeXt-WAM for latent-action, efficient, larger-scale, JEPA-style, and World-Action-Model exploration.The authors present these as new baselines built from the expanded codebase.Source: §5 · p. 10

Design implications

• The conclusion identifies VLM–policy interaction, proprioception fusion, and temporal action modeling as central design choices in VLA performance.Soft VLM–policy coupling and VLM-side proprioception conditioning show that where information enters the pipeline matters as much as which information is used.Source: §6 · p. 10
• Structured sequence modeling, including frequency-domain objectives, transfers time-series learning ideas effectively to robotics, while world modeling improves performance at notable computational cost.The conclusion therefore emphasizes efficiency-aware design when adopting richer objectives.Source: §6 · p. 10

Additional baselines

• Compact models remain competitive, while larger VLM backbones, latent-action pretraining, latent predictive visual objectives, and world action modeling provide complementary gains.These additional baselines extend the exploration across model scales and several forms of latent or world-action modeling.Source: §6 · p. 10

Future directions

• The released unified, lightweight framework is intended to replace ad-hoc VLA variations with controlled design-space studies and shared progress.The authors position the codebase as infrastructure for systematic exploration rather than another standalone architecture.Source: §6 · p. 10
• Future work should address more diverse embodiments, longer-horizon reasoning, extensive mid-training, and richer world-interaction objectives.These directions extend the systematic perspective beyond the current exploration.Source: §6 · p. 10

• The work targets improved Vision-Language-Action models for robotic control and the development of more capable embodied agents.Source: Impact Statement · p. 10
• The authors consider the work’s societal implications part of established discussions in the field and not requiring special emphasis.Source: Impact Statement · p. 11

A. More Experimental Results

• Additional demonstrations cover LIBERO, LIBERO-plus, and real-world experimental settings.The supplementary results direct readers to Figures 10, 11, and 9 and accompanying video demonstrations.Source: §A.1 · p. 17
• The supplementary material provides video demonstrations of the reported experiments.The videos accompany demonstrations from LIBERO, LIBERO-plus, and real-world settings.Source: §A.1 · p. 17

A.1. Qualitative Experiments

• Real-world qualitative experiments illustrate the method’s behavior on practical tasks.Figure 9 presents qualitative experiments in real-world tasks.Source: Figure 9 · p. 17
• Together, the qualitative figures organize the supplementary evidence by real-world tasks, LIBERO suites, and LIBERO-plus perturbations.Figures 9–11 correspond respectively to real-world tasks, the four LIBERO suites, and seven LIBERO-plus perturbation types.Source: Figure 9 · p. 17, Figure 10 · p. 17, Figure 11 · p. 18
• The LIBERO qualitative results span all four benchmark suites.Figure 10 shows qualitative experiments across the four suites of LIBERO.Source: Figure 10 · p. 17
• The LIBERO-plus demonstrations cover seven perturbation types applied to the same task.Figure 11 visualizes qualitative experiments under the seven perturbations.Source: Figure 11 · p. 18

Cross-suite ablations

• Results in the Object, Goal, and Long suites are consistent with the Spatial-suite experiments on LIBERO and LIBERO-plus, demonstrating recipe robustness.These results are reported in Tables 6 and 7 and compared with the Spatial-suite results in Table 1.Source: §A.2 · p. 18
• The additional LIBERO and LIBERO-plus ablations follow the same experimental setup as the main experiments in Table 1.Source: §A.2 · p. 18

Real-world ablation

• The real-world ablation evaluates the clean table task using the setup from Table 4 in the main paper.Source: §A.2 · p. 18

A.3. Detailed Experimental Settings

• The final model uses the same training parameters across all four benchmark suites, supporting the claimed robustness of the recipe.Detailed hyperparameters and benchmark and real-world ablation results are provided in Tables 6–8, although the supplied passages do not report their cell values.Source: §A.3 · p. 18, Table 8 · p. 19, Table 6 · p. 19

B. Revisiting Robot Learning and VLA Models

• The work focuses on general-purpose robotic manipulation through imitation learning because diverse goals and objects make explicit rewards and task priors difficult to design.Earlier manipulation research often targeted specific grasping or non-prehensive tasks with clearer priors, whereas general-purpose manipulation requires learning complex skills from expert demonstrations.Source: B. Revisiting Robot Learning and VLA… · p. 20
• Robot-learning tasks are commonly divided into locomotion and manipulation, with locomotion often using reinforcement learning and manipulation increasingly addressing broad interactive skills.Locomotion controls robot-base stability and mobility, while manipulation controls an arm or whole body to execute interactive tasks.Source: B. Revisiting Robot Learning and VLA… · p. 20
• Manipulation policies include standard action policies that directly predict actions and video action policies that jointly predict future videos and actions.Video-action methods claim that world-modeling abilities can improve task understanding and action accuracy.Source: B. Revisiting Robot Learning and VLA… · p. 20
• VLA research began with RT-2 and expanded toward 3D information, intermediate predictions, post-training optimization, latent actions, lightweight models, and domain-specific systems.Despite these emphases, most models use pretrained LLMs or VLMs to process visual observations and language instructions into action-relevant policy representations.Source: B. Revisiting Robot Learning and VLA… · p. 20