agibot-world/AgiBotWorldChallenge-2025

AgiBot World Challenge 2025 - Datasets

Dear participants,

We are excited to announce that the datasets for both tracks of AgiBot World Challenge 2025 have been updated.

---

Changelog

Previous versions remain available in the branch version name.

• v1.0.1 (27-06-2025): Fixed an Issue with the annotations in the simulation data of the "open drawer and store items" task.
• v1.0.0 (25-06-2025): Initial version. Released datasets for both tracks of AgiBot World Challenge 2025.

Track 1：Manipulation

We have specifically collected data for 10 distinct tasks for this competition, with hundreds of trajectories per task. Utilizing an advanced data collection approach - Adversarial Data Collection (ADC), we've incorporated dynamic disturbances to significantly enhance the information density and diversity of each trajectory. This approach not only reduces post-training data requirements and model training costs but also effectively strengthens the model's generalization capabilities and robustness.

The simulation datasets for 10 tasks are provided in the Manipulation-SimData folder. To facilitate local simulation evaluation for participants, we have open-sourced a set of simulation assets (path: https://huggingface.co/datasets/agibot-world/GenieSimAssets).

Each real-robot dataset has one or two corresponding files, and here is the mapping between task names and their respective dataset IDs:

Task Name	Real Robot Dataset ID
Heat the food in the microwave	881
Open drawer and store items	949, 1019
Pack in the supermarket	1352, 1418
Stamp the seal	1458
Pack washing detergent from conveyor	1645
Clear the countertop waste	1957
Pickup items from the freezer	1918, 2055
Restock supermarket items	1967
Make a sandwich	1969
Clear table in the restaurant	1968

---

Track 2：World Model

We've designed a series of challenging tasks covering scenarios including kitchen environments, workbenches, dining tables, and bathroom settings, encompassing diverse robot-object interactions (e.g., collisions, grasping, placement, and dragging maneuvers), to thoroughly evaluate models' generative capabilities.

This track offers a comprehensive dataset consisting of training, validation, and testing sets:

Dataset Structure

• Training Set
  The training set includes over 30,000 premium trajectories selected from 10 representative tasks in the AgiBot World Dataset, providing ample material for model training.

• Validation Set
  The validation set contains 30 carefully chosen samples to support model verification and optimization.

• Testing Set
  The testing set includes 30 no-public samples, covering both seen and unseen scenarios in the training set, mixed with expert demonstrations and imperfect trajectories, aiming to assess models' generalization and robustness comprehensively.

DATASET_ROOT/
├── train/
│   ├── 367-648961-000/
│   │   ├── head_color.mp4
│   │   ├── head_extrinsic_params_aligned.json
│   │   ├── head_intrinsic_params.json
│   │   └── proprio_stats.h5
│   ├── 367-648961-001/
│   │   ├── head_color.mp4
│   │   ├── head_extrinsic_params_aligned.json
│   │   ├── head_intrinsic_params.json
│   │   └── proprio_stats.h5
│   ├── {task_id}-{episode_id}-{step_id}/
│   │   ├── head_color.mp4
│   │   ├── head_extrinsic_params_aligned.json
│   │   ├── head_intrinsic_params.json
│   │   └── proprio_stats.h5
│   └── ...
├── val/
│   ├── 367-649524-000/
│   │   ├── head_color.mp4
│   │   ├── head_extrinsic_params_aligned.json
│   │   ├── head_intrinsic_params.json
│   │   └── proprio_stats.h5
│   └── ...
└── test/
    ├── {task_id}-{episode_id}-{step_id}/
    │   ├── frame.png
    │   ├── head_color.mp4 (NOT disclosed to participants)
    │   ├── head_extrinsic_params_aligned.json
    │   ├── head_intrinsic_params.json
    │   └── proprio_stats.h5
    └── ...

Provided Data Includes

• EEF poses
• Joint angles
• Camera intrinsics/extrinsics
• ......

→ Enabling participants to fully utilize physical and visual information.

---

We look forward to seeing your innovative solutions in the challenge!