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

1. Introduction

• Real-robot data supports VLA generalization but is resource-intensive to collect, while existing simulated datasets cover narrow skills and mostly rigid objects and rarely validate large-scale VLA pre-training.Real-data collection requires skilled teleoperators, specialized hardware, and extensive labor; prior simulated datasets also require nontrivial human operation. The paper frames whether sufficiently diverse, high-fidelity synthetic data can match the strongest real-world pre-training data as an open question.Source: §1 · p. 2
• InternData-A1 contains 630k trajectories and 7,433 hours spanning 4 embodiments, 18 skills, 70 tasks, and 227 scenes, with rigid, articulated, deformable, and fluid-object interactions.Its autonomous, fully decoupled pipeline separates asset specification, skill policies, task composition, and rendering; framework optimizations produce 209.7 hours of robot data per day on 8 RTX 4090 GPUs at below $0.003 per episode, with minimal manual tuning.Source: §1 · p. 2
• A π0 model pre-trained exclusively on InternData-A1 closely matches the official π0 across 49 simulation, 5 real-world, and 4 long-horizon dexterous tasks, and outperforms models trained on existing open-source datasets.The official π0 was trained on the closed-source π-dataset; the authors present this as evidence that large-scale simulation-only data can match the strongest real-world data for VLA pre-training.Source: §1 · pp. 2–3
• Ten selected simulated tasks achieve direct sim-to-real transfer with an average success rate exceeding 50%.On four representative tasks under well-aligned sim-to-real settings, fewer than 1,600 simulated samples can match the performance of 200 real samples.Source: §1 · p. 3
• InternData-A1 and its generation pipeline are open-sourced and reproducible in simulation to broaden access to embodied AI research and data-scaling studies.Source: §1 · p. 3

Large-scale Robot Datasets

• InternData-A1 contrasts with prior simulation datasets through broad task diversity, heterogeneous embodiments, photorealistic rendering, and long-horizon multi-skill trajectories.Earlier simulation datasets support controllable variation and augmentation but require substantial manual effort or mainly target picking; RoboTwin 2.0 (Chen et al., 2025a) expanded to bimanual manipulation. Real-world data offers high-fidelity demonstrations but is costly to scale and often has limited embodiment or scene diversity or variable quality.Source: §2 · p. 3

VLA Models

• VLA pre-training approaches use real-only, simulation-only, or hybrid data; each faces limits in open-world generalization, skill coverage, or grounding and planning.Real-data–only models can perform strongly in-domain but may remain tied to training environments; simulation-only models often cover narrow skill families, while hybrid approaches combine simulation with demonstrations or web-scale VQA data.Source: §2 · p. 3
• InternData-A1 is used to strengthen action priors and generalization, with downstream performance reported on par with state-of-the-art real data despite simulation gaps.The study adopts the π0 model and its underlying π-dataset as its primary baselines.Source: §2 · p. 3

• The statistics cover embodiments, tasks, and assets, with overall scale and diversity visualized in Figure 2.Source: §3 · p. 4

Embodiments

• The four embodiments are AgiBot Genie-1 (7.9%), Franka Emika Panda (23.3%), AgileX Split Aloha with Piper-100 arms (30.8%), and ARX Lift-2 with R5a arms (37.8%).Franka is fixed on a table for tabletop manipulation; the two dual-arm systems support bimanual manipulation, including sequential and parallel behaviors.Source: §3 · p. 4

Tasks

• Seventy tasks span rigid, articulated, deformable, and fluid manipulation, each defined by context, atomic-skill composition, and action-space constraints.Eighteen long-horizon tasks involve at least three sequential skills, totaling 124,789 trajectories and 141,421,619 frames; 56 tasks contain 1,000–10,000 trajectories.Source: §3 · p. 4

Assets

• The inventory includes 3,185 rigid objects in 107 categories, 321 articulated objects in 14 categories, 20 scanned garments, and 227 indoor scenes.The indoor scenes include kitchens, study rooms, dining rooms, and living rooms.Source: §3 · pp. 4–5

4.1. Environment Construction

• Annotated robot, scene, and object assets support high-fidelity manipulation across rigid, articulated, deformable, and fluid categories.Scenes include manipulation-area metadata; object annotations include physical properties and interaction features, with deformables and fluids simulated using specialized methods.Source: §4.1 · pp. 5–6

4.2. Skill Composition

• Modular skills map object and robot states plus constraints to waypoints, enabling sequential or parallel long-horizon tasks with only spatial-range adjustments.Users configure skills through command templates and select the arm for each skill; waypoint representations decouple high-level skill logic from low-level motion execution.Source: §4.2 · p. 6

4.3. Domain Randomization

• Visual and trajectory diversity comes from camera perturbations, 174 randomized environment maps, category-based object replacement, spatial sampling, and randomized contact selection.Grasp poses are selected randomly from the top 40 high-confidence candidates after filtering; articulated and deformable-object contact points are sampled within expanded contact neighborhoods.Source: §4.3 · p. 6

4.4. Generation & Storage

• CuRobo interpolates dense joint-space actions between skill waypoints, and episodes record language, multi-view images, camera parameters, robot states, and action labels in LeRobot format.Configuration options also allow recording depth maps, grounding annotations, and bounding boxes.Source: §4.4 · p. 6

4.5. Framework Optimization

• 2–3× end-to-end performance improvement over the baseline comes from pipelined planning and rendering, dynamic resource scheduling, stacked rendering, and cluster-balancing mechanisms.The design addresses falling planning success as task complexity rises and the poor hardware utilization caused by serially executing CPU-bound planning and GPU-based rendering.Source: §4.5 · p. 7

5.1. Comparison with 𝜋-dataset

• The comparison uses π0 architecture and identical downstream fine-tuning; InternData-A1-only pretraining starts from Paligemma weights with a randomly initialized action expert.The open-source-dataset comparison pretrains each π0 model for 500k iterations and evaluates on 49 simulated and two real-world tasks.Source: §5.1 · p. 7, §5.2 · p. 9
• Across 49 simulation tasks, a π0 model pretrained on InternData-A1 outperformed the official π0 model by 5% in Easy settings and 6.5% in Hard settings, respectively.Source: Table 2 · p. 8
• Across five regular real-world tasks, InternData-A1 outperforms 𝜋-dataset by 6.2% across in-distribution and out-of-distribution tasks.The tasks span two embodiments and cover articulation, pick-and-place, human–robot interaction, and tool use.Source: §5.1 · p. 9
• On four long-horizon dexterous tasks, InternData-A1 performs comparably to 𝜋-dataset despite novel objects, skills, and an unseen ARX AC One embodiment.The tasks include garment folding, industrial-part sorting, bottle-cap unscrewing, and bag zipping.Source: §5.1 · p. 9, Figure 5 · p. 9

5.2. Benchmarking against open-source datasets

• In real-robot evaluations, InternData-A1 improves average performance over RoboCasa by 57.7%, although RoboCasa trails it by only 10% in simulation.The authors attribute the real-world gain to InternData-A1’s photorealistic rendering and abundant data.Source: §5.2 · pp. 9–10, Table 3 · p. 10

6. Data Analysis

• 200 simulated episodes matched 200 real episodes on basic tasks, while about 1,600 were needed for complex tasks, keeping equivalent-performance ratios within 8:1 and sometimes near 1:1.Coarse visual alignment was sufficient when camera views and joint action spaces were closely aligned; the authors attribute robustness to photorealistic rendering and extensive domain randomization.Source: Figure 6 · p. 10, §6 · pp. 10–11
• With 500 simulated episodes, success was 50% on Make Sandwich and Pack, 63% on Close Box, 87% on Close Microwave, 60% on Sweep, and 57% on Handover.Together with the four Figure 6 tasks, ten representative tasks out of seventy achieved high sim-to-real success without real data.Source: Figure 7 · p. 11, §6 · p. 11
• The authors describe InternData-A1 as the first to demonstrate direct VLA sim-to-real transfer for diverse, complex tasks beyond single-pick tasks, including multi-skill, articulation, bimanual, and logistics Flip Package tasks.They contrast this with GraspVLA (Deng et al., 2024), which demonstrated single-pick transfer, and note that Helix (Cui et al., 2025) previously realized Flip Package using purely real data.Source: §6 · p. 11

6.1. Data Component Ablation

• The full-data model achieved the highest RoboTwin 2.0 success rates, and removing any of PnP, Art, Base, or Long reduced performance.Removing Base or Long caused larger declines than removing PnP, and excluding Art also caused a larger drop than excluding PnP; the authors hypothesize trajectory diversity is central to effective pre-training. The findings suggest datasets limited to pick-and-place tasks, such as those in Chen et al. (2025b) and Deng et al. (2024), are insufficient.Source: §6.1 · p. 12, Table 4 · p. 12

Conclusion

• Synthetic-only pre-training on InternData-A1 matches the performance of the leading VLA model.The paper presents this as the first demonstration of this result.Source: §7 · p. 12
• Open-sourcing InternData-A1 and its generation pipeline aims to lower barriers to large-scale robotic data access for embodied AI.Source: §7 · p. 12

Limitation

• Physics simulator limitations make highly dexterous tasks, such as tying shoelaces or threading a needle, challenging to simulate.Future work will expand task diversity and dexterity to further establish large-scale simulation data as a cornerstone for VLA-model advancement.Source: §7 · p. 12

• The dataset spans 4 embodiments and 70 tasks, with 637,498 trajectories, 401,430,981 frames, and 7,433.91 hours of interaction data.Source: A. Detailed Data Statistics · p. 18
• Across 637,498 total trajectories, the dataset’s four categories are Articulation (74,415; 11.67%), Long-horizon (138,782; 21.77%), Base (229,168; 35.95%), and Pick and Place (195,133; 30.61%).Source: A. Detailed Data Statistics · p. 18

B. Detailed Data Synthesis

• The appendix presents an example task configuration and explains its components in detail.Source: B. Detailed Data Synthesis · p. 18

B.1. Environment Construction

• The configuration selects a dining-room layout, Agilex Split Aloha robot, and its motion planner through the robot_file.Agilex Split Aloha is one of the four embodiments.Source: §B.1 · p. 18
• Each task asset is automatically annotated with gravity, collision, and grasp parameters, while initial object poses are assigned using category-level canonical definitions.Source: §B.1 · p. 18

B.2. Skill Composition

• Tasks can chain skills sequentially or run them in parallel, including coordinated gripper actions, to express diverse task requirements.A task may chain pick, goto_pose, gripper_action, home, and place skills.Source: §B.2 · p. 18
• Task-level placement constraints align object axes for accurate insertion, while ratio ranges specify the target insertion layer.Source: §B.2 · pp. 18–19
• Users can tune grasp-pose filters and grasp or placement offsets to improve stability and avoid unsafe motions.Source: §B.2 · p. 19

B.3. Domain Randomization

• Randomized lighting, camera poses, room types, and same-category object replacements diversify visual conditions and scene contents.Camera extrinsics can vary by up to 5° in rotation and 5 cm in translation.Source: §B.3 · p. 19
• Episode diversity also comes from randomized object and robot initial poses, varied home joint states, and random selection among valid grasp and placement poses.The framework retains poses satisfying loose filtering ranges, then selects a final pose randomly.Source: §B.3 · p. 19

C. Policy Training Details

• 680k iterations on InternData-A1 across 32 A100 GPUs pretrain a new π0 model, closely matching the official π0 checkpoint’s 700k training iterations.The model is initialized with Paligemma weights and a scratch-initialized action expert.Source: C. Policy Training Details · p. 19
• Post-training starts from the 680k π0(InternData-A1) checkpoint and runs 30k iterations for regular and sim-to-real tasks, or 100k for dexterous tasks.Key training hyperparameters are summarized in Table 6.Source: C. Policy Training Details · p. 19, Table 6 · p. 19

D. Real-World Experiments

• Each task uses 15 evaluation settings with two trials each, yielding 30 rollouts; reported success rate is the average across rollouts.The evaluated checkpoint is taken after 30k post-training iterations.Source: D. Real-World Experiments · p. 19

D.1. Real Task Description

• Place Markpen tests Genie-1 right-arm pick-and-place, succeeding only when a black marker is placed precisely and fully inside a pen holder.Source: §D.1 · p. 20
• Pass Bottle tests Genie-1’s right-arm pickup, upright lift, and handoff of a black tea bottle; release is allowed only when a person presents a hand.Success requires transfer to the person and corresponding gripper release.Source: §D.1 · p. 20
• Heat Sandwich tests ARX Lift-2 articulated-object operation: its left arm opens and closes the oven while its right arm inserts the sandwich plate.Success requires correct plate insertion and a fully closed oven door.Source: §D.1 · p. 20
• Sort Rubbish tests ARX Lift-2’s right-arm sorting of recyclable waste into the right bin and non-recyclables into the left bin.Success requires every item to be fully and correctly sorted.Source: §D.1 · p. 20
• Sweep Trash requires ARX Lift-2 to use both arms to sweep crumpled paper into a dustpan, empty it into the left bin, then release both grippers and return home.Every step must be completed successfully.Source: §D.1 · p. 20
• Sort Parts tests ARX Lift-2’s two-arm sorting of four types of small industrial components into designated containers, with each arm handling two categories.Success requires all components in their correct containers.Source: §D.1 · p. 20
• ARX AC One’s Unscrew Cap task requires its left arm to position the tea bottle and its right arm to remove the cap; Fold Cloths requires both hands to fold cloth correctly.Zip Bag requires the robot’s left arm to open a bag, place designated objects inside, and zip it fully and correctly.Source: §D.1 · p. 20

D.2. Sim-to-real Task Description

• Flip Package requires grasping, flipping, and returning a package before scanning its QR code; Instructional Pick requires selecting the commanded object, while Wipe Stain requires removing all stains.The package flip uses the right arm, followed by a left-arm grasp and robot-mounted camera scan; the wipe uses the left arm on one or two stain clusters.Source: §D.2 · p. 20
• Sandwich requires assembling bread and beef neatly, while Box and Microwave require closing lids and Pack requires placing all objects inside a box.Sandwich uses alternating arms; Box uses both arms, while Microwave uses the right arm.Source: §D.2 · pp. 20–21
• Handover transfers a long-shaped object from the left arm to the right arm for placement in a box; Sort Rubbish and Sweep repeat tasks described earlier.Handover succeeds only when the object reaches the box; Sort Rubbish and Sweep are not further specified here.Source: §D.2 · pp. 20–21