InternData-A1: Pioneering High-Fidelity Synthetic Data for Pre-training Generalist Policy

TL;DR

This work shows that large-scale, high-fidelity synthetic data can match the best real-robot pre-training for Vision-Language-Action models. The authors present InternData-A1, a decoupled, autonomous data-synthesis pipeline producing 630k trajectories (7,433 hours) across 4 embodiments, 70 tasks, and 227 scenes with photorealistic rendering and broad skill composition. A $\pi_0$ model pre-trained solely on InternData-A1 achieves performance on par with or better than the official $\pi_0$ across 49 simulation tasks and 9 real-world tasks, with direct sim-to-real transfer observed on multiple tasks and favorable comparisons to open-source datasets. The dataset and pipeline are open-source, underscoring the potential of large-scale simulation to reduce barriers to scalable embodied AI data while highlighting remaining challenges in highly dexterous manipulation.

Abstract

Recent works explore how real and synthetic data contribute to Vision-Language-Action (VLA) models' generalization. While current VLA models have shown the strong effectiveness of large-scale real-robot pre-training, synthetic data has not previously demonstrated comparable capability at scale. This paper provides the first evidence that synthetic data alone can match the performance of the strongest $π$-dataset in pre-training a VLA model, revealing the substantial value of large-scale simulation. The resulting model also exhibits surprisingly zero-shot sim-to-real transfer on several challenging tasks. Our synthetic dataset, InternData-A1, contains over 630k trajectories and 7,433 hours across 4 embodiments, 18 skills, 70 tasks, and 227 scenes, covering rigid, articulated, deformable, and fluid-object manipulation. It is generated through a highly autonomous, fully decoupled, and compositional simulation pipeline that enables long-horizon skill composition, flexible task assembly, and heterogeneous embodiments with minimal manual tuning. Using the same architecture as $π_0$, we pre-train a model entirely on InternData-A1 and find that it matches the official $π_0$ across 49 simulation tasks, 5 real-world tasks, and 4 long-horizon dexterous tasks. We release the dataset and will open-source the generation pipeline to broaden access to large-scale robotic data and to lower the barrier to scalable data creation for embodied AI research.

Figures (7)

• Figure 1: InternData-A1 pioneers a large-scale, high-fidelity synthetic dataset with physically faithful, photorealistic rendering, diverse object domains (rigid, articulated, fluid, and deformable), extensive multi-skill tasks, and broad cross-embodiment coverage.
• Figure 2: Data Statistics. InternData-A1 provides 4 single or dual-arm embodiments, 70 diverse tasks, 3185 rigid objects, 321 aritculation objects, 20 garments, and 227 rooms. All these elements consist 630k episodes, 401.4M frames and 7433.9 hours.
• Figure 3: Data synthesis pipeline of InternData-A1. It consists of four stages: (1) environment construction with selected embodiments, scenes, and objects; (2) task composition using modular atomic skills invoked via simple configuration commands; (3) domain randomization over layouts, object poses, lighting, etc.; and (4) trajectory generation, where CuRobo curobo_report23 interpolates dense joint actions, validates them through physics simulation, and renders only successful trajectories into the LeRobot format.
• Figure 4: Real-world setup. We evaluate the model pre-trained on InternData-A1 against the $\pi$-dataset baseline on a suite of nine real-world tasks spanning three robots.
• Figure 5: Comparison to $\pi$-dataset in real-world tasks. InternData-A1 achieves performance comparable to $\pi$-dataset across 9 real-world tasks, including 4 dexterous ones, demonstrating its strong pre-training capability.