Datasets:

chen-gao
/

dualsim

license: cc-by-nc-4.0
task_categories:
  - robotics
tags:
  - robotics
  - manipulation
  - lerobot
  - agibot-world
  - dual-arm
  - simulation
  - world-model
pretty_name: DualSim (AgiBotWorld Challenge 2026)
configs:
  - config_name: eval
    data_files: eval/AgiBotWorldChallenge-2026-eval-dualsim/_lerobot/data/**/*.parquet

DualSim

Data package for the AgiBotWorld Challenge 2026 dual-arm simulation track.

Layout

dualsim/
├── eval/                       # dualsim-specific evaluation set (~321 MB)
│   └── AgiBotWorldChallenge-2026-eval-dualsim/
│       ├── _lerobot/           # LeRobot v2.1 dataset (meta / data / videos / sim)
│       ├── info_dataset/       # per-episode inputs (frame.png, proprio_stats.h5, camera params, sim/)
│       ├── gt_dataset/         # per-episode ground truth
│       └── list.csv            # episode index (task, frames, fps, duration)
└── train/
    └── AgiBotWorld2026/        # training data (mirror of agibot-world/AgiBotWorld2026)
        └── RichInteraction/Home/...

Eval set summary

Format: LeRobot v2.1, robot type g2a
11 episodes · 773 frames · 1 task · 5 fps
Observation: observation.images.top_head (480×640×3 video) + observation.state (42-dim)
11 evaluation episodes across tasks 367 / 392 / 511 / 515 / 525 / 558 / 573 (see list.csv)

Training data

train/AgiBotWorld2026/ mirrors the public dataset agibot-world/AgiBotWorld2026 (RichInteraction/Home split). If you already have it locally, you can symlink it in instead of downloading again.

Updating

This repo is synced with hf upload, which only transfers files that changed. To push local edits back to the Hub, run ./sync_dualsim.sh from the dataset root (requires hf auth login).