InternData-A1

A large-scale simulated synthetic dataset from Shanghai AI Lab, built for pretraining general-purpose robot policies.

InternData-A1 is a synthetic robot-manipulation dataset released in November 2025 by the Shanghai Artificial Intelligence Laboratory (the InternRobotics team, with participation from Peking University), generated entirely in simulation: over 630,000 trajectories totaling 7,433 hours, covering 4 robot embodiments (ARX Lift-2, AgileX Split ALOHA, A2D, and Franka), 70 tasks, and 227 scenes, including manipulation of rigid, articulated, and deformable objects as well as fluids. The pipeline decouples skills, tasks, and embodiments so they can be freely recombined, uses cuRobo for motion planning, and applies domain randomization (randomly varying things like camera viewpoint, lighting, and object placement so the model doesn't depend on any one fixed appearance). The paper's conclusion: pretraining π0 on this synthetic data alone matches the performance of the official π0, which was pretrained on π's real-robot data, across 49 simulated tasks, 5 real-robot tasks, and 4 long-horizon dexterous tasks. The data is released on Hugging Face under CC BY-NC-SA 4.0.

ExampleThe paper reports that for some tasks, fewer than 1,600 simulated samples match the effect of 200 real-robot samples; for 10 of the 70 tasks, the policy achieves strong real-robot success without using any real-robot data at all.

Also called

InternData A1

Related

Synthetic Data · Simulation Data · π0 · InternVLA (Shanghai AI Laboratory) · cuRobo (NVIDIA GPU-accelerated motion planning) · Domain Randomization

Sources

InternData-A1: Pioneering High-Fidelity Synthetic Data for Pre-training Generalist Policy (arXiv 2511.16651)
InternData-A1 数据集页面（Hugging Face） (Chinese)

As of

2026-01