InternVLA-A1: Unifying Understanding, Generation and Action for Robotic Manipulation

Abstract: Prevalent Vision-Language-Action (VLA) models are typically built upon Multimodal LLMs (MLLMs) and demonstrate exceptional proficiency in semantic understanding, but they inherently lack the capability to deduce physical world dynamics. Consequently, recent approaches have shifted toward World Models, typically formulated via video prediction; however, these methods often suffer from a lack of semantic grounding and exhibit brittleness when handling prediction errors. To synergize semantic understanding with dynamic predictive capabilities, we present InternVLA-A1. This model employs a unified Mixture-of-Transformers architecture, coordinating three experts for scene understanding, visual foresight generation, and action execution. These components interact seamlessly through a unified masked self-attention mechanism. Building upon InternVL3 and Qwen3-VL, we instantiate InternVLA-A1 at 2B and 3B parameter scales. We pre-train these models on hybrid synthetic-real datasets spanning InternData-A1 and Agibot-World, covering over 533M frames. This hybrid training strategy effectively harnesses the diversity of synthetic simulation data while minimizing the sim-to-real gap. We evaluated InternVLA-A1 across 12 real-world robotic tasks and simulation benchmark. It significantly outperforms leading models like pi0 and GR00T N1.5, achieving a 14.5\% improvement in daily tasks and a 40\%-73.3\% boost in dynamic settings, such as conveyor belt sorting.

• The paper introduces an integrated model that fuses semantic reasoning, visual foresight, and action generation to overcome limitations in dynamic robotic tasks.
• It employs a Mixture-of-Transformers with dedicated experts and a hierarchical data pyramid to enhance sim-to-real transfer and multi-agent coordination.
• Experimental results demonstrate significant improvements over baselines, with up to 93.3% success in dynamic and coordinated manipulation scenarios.

Unified Vision-Language-Action Model for Robust Robotic Manipulation: An Expert Review of InternVLA-A1 (2601.02456)

Introduction

InternVLA-A1 represents a significant advance in generalist robot policy learning, targeting the limitations of conventional Vision-Language-Action (VLA) architectures. Previous state-of-the-art VLA models, often based on Multimodal LLMs (MLLMs), have excelled at semantic task understanding but remain incapable of reasoning about physical world dynamics, essential for effective robotic action in dynamic environments. InternVLA-A1 addresses these gaps by unifying semantic reasoning, visual foresight (predictive dynamics), and action generation into a single, scalable model. This integration, realized via a Mixture-of-Transformers (MoT) architecture with three specialized experts, is further buttressed by a hierarchical data pyramid leveraging large-scale synthetic, real-world, and fine-tuning datasets.

Figure 1: InternVLA-A1’s architecture tightly couples semantic understanding and predictive dynamics, yielding robust performance in highly dynamic scenarios.

Model Architecture

The core innovation of InternVLA-A1 lies in its tripartite expert architecture realized through the Mixture-of-Transformers framework:

• Understanding Expert: Utilizes open-source MLLM backbones (InternVL3, Qwen3-VL) to encode multimodal sensory inputs (multi-view images and text instructions) into unified semantic tokens.
• Generation Expert: Deploys a VAE-based visual tokenizer (Cosmos CI) enabling high-fidelity latent encoding. It predicts future scene states in latent space, crucial for anticipatory control, and is optimized for real-time inference via spatial compression and efficient architectural design.
• Action Expert: Consumes context and predictive dynamics, employing a Flow Matching objective for continuous multimodal action generation. This yields high quality, diverse control outputs applicable to multiple robot embodiments.

A blockwise attention mask enforces an explicit information flow, ensuring each expert attends only to relevant upstream representations. All components are jointly optimized for visual foresight and action prediction via weighted loss terms, facilitating robust causal modeling across domain boundaries.

Figure 2: InternVLA-A1’s framework comprises dedicated experts for understanding, generation, and action, connected via masked attention and flow matching for robust, anticipatory manipulation.

Hierarchical Data Pyramid and Training Strategy

InternVLA-A1 achieves strong out-of-distribution generalization by pre-training on a hierarchical data pyramid:

• Synthetic Data (InternData-A1): High-diversity simulated trajectories, supporting skill composition and heterogeneous embodiment. Demonstrated to close the performance gap with the largest closed real-world datasets.
• Real-World Demonstrations (Agibot-World): Over one million real trajectories spanning 217 tasks, mitigating sim-to-real discrepancies and improving physical reasoning.
• Specialized Fine-Tuning Data: High-quality teleoperated episodes for target domains, enabling domain adaptation for specific deployment environments.

This data recipe allows InternVLA-A1 to scale skill acquisition efficiently, balance diversity and realism, and achieve superior sim-to-real transfer compared to prior art.

Figure 3: Hierarchical data pyramid—synthetic, real-world, and specialized datasets each underpin a phase of InternVLA-A1’s training.

Experimental Evaluation

General-Purpose Manipulation Tasks

InternVLA-A1 was evaluated against leading baselines ($\pi_0$, GR00T N1.5) on ten diverse real-world manipulation tasks. The 2B variant already outperforms $\pi_0$ (3.3B) despite lower parameter count, and the 3B variant achieves a 14.5% higher average success rate (75.1% vs. 60.6%). Notably, task robustness is evident in both long-horizon (e.g., Make Sandwich: 93.3% vs. 66.7%) and dexterous tasks (e.g., Zip Bag: 73.3% vs. 40.0% for $\pi_0$).

Figure 4: Experimental setting for general-purpose tasks spanning dexterous and regular manipulation.

High-Dynamics/Coordination Tasks

InternVLA-A1 demonstrates marked superiority in dynamic environments. On Express Sorting and In-motion Ingredient Picking (requiring reasoning about future scene states and multi-robot coordination), it delivers 40–73.3% absolute improvements over baselines. The 3B model attains up to 93.3% success in multi-agent tasks, clearly exceeding prior model capabilities for anticipatory and coordinated manipulation.

Figure 5: Experimental protocol for dynamic manipulation benchmarks—Express Sorting and coordinated Ingredient Picking.

Figure 6: InternVLA-A1 outperforms prior models by wide margins on dynamic task families, evidencing exceptional stability.

Simulation Benchmark

On the RoboTwin 2.0 benchmark (bimanual tasks with domain randomization), InternVLA-A1 consistently exceeds $\pi_0$ by 10.5% (Easy) and 5.6% (Hard) in task success, showing effective adaptation to challenging simulated conditions.

Ablation Studies

• Pre-training: Eliminating pre-training causes a dramatic 51.6% performance regression (down to 25.4% success rate), confirming its role as a critical inductive prior.
• Generation Expert: Ablating the generation expert consistently hurts performance (−19.4% overall), particularly in dynamic tasks, establishing its necessity for causal reasoning and robustness.
• Data Source Impact: Synthetic data (InternData-A1) alone rivals real data (Agibot-World) in pre-training. Mixed pre-training (synthetic + real) synergizes strengths, improving real-world task success (e.g., +20% in Sort Parts), with domain-specific variation in sim-to-sim and sim-to-real transfer.

  Figure 7: Pre-training ablation—internally pre-trained models greatly stabilize optimization and improve downstream generalization.

  

  Figure 8: Removing the generation expert substantially degrades real-world success across nearly all task categories.

  

  Figure 9: Hybrid pre-training (synthetic + real) integrates strengths of each, maximizing downstream performance and sim-to-real adaptation.

Implications and Future Directions

InternVLA-A1 establishes robust, dynamic-aware manipulation as a tractable problem in end-to-end policy learning by fusing semantic reasoning and predictive dynamics. The architecture’s modularity suggests extensibility to additional sensor or policy experts, joint VQA training for improved instruction following, and higher fidelity modeling of visual foresight. Practically, its hierarchical data recipe demonstrates scalable skill bootstrapping and domain transfer for real-world robotic platforms. Theoretically, InternVLA-A1 exemplifies how joint optimization of multimodal transformers for understanding, generation, and action can achieve superior scene and action generalization.

Remaining limitations include incomplete joint training for semantic understanding, which degrades multi-turn instruction comprehension, and necessary compromises in predictive frame resolution for real-time inference. Future research should focus on model scaling, instruction-following augmentation via VQA datasets, and improved generative fidelity for physical prediction.

Conclusion

InternVLA-A1 (2601.02456) delivers a unified, dynamics-aware framework for robotic manipulation by coordinating semantic context, visual dynamics prediction, and action synthesis. Through architectural innovations and a hierarchical pre-training strategy, it outperforms prior state-of-the-art VLA models across diverse static and dynamic manipulation tasks, establishing new benchmarks for robustness, adaptability, and sim-to-real transfer. The work provides a foundation for scalable, generalist robot policies and identifies clear directions for further refinement, especially in multimodal semantic reasoning and real-time generative modeling.