The Dataset Viewer has been disabled on this dataset.

VI AMO v2

VI AMO pairs robot demonstration videos with temporally aligned action and mission annotations for hierarchical intent prediction. VI AMO v2 contains 6,600 episodes, 13,653 actions, and 9,702 missions from eleven robot data sources. This main release contains the frozen 2026-10-14 annotations used for the arXiv paper's offline training and benchmark evaluation.

Split Episodes Actions Missions
Train 5,500 11,481 8,129
Validation 550 1,081 792
Test 550 1,091 781
Total 6,600 13,653 9,702

File Layout

vi_amo/
  README.md
  train.jsonl
  val.jsonl
  test.jsonl
  meta/
    info.json
    training_run.json
    checksums.sha256
  videos/
    chunk-000/
      episode_000000/
        front.mp4
        wrist.mp4
      ...
    ...

The dataset contains 15,735 MP4 files, with one file per episode and camera. Camera availability varies by source. Each episode retains its original frame rate and image resolution. Existing episode videos retain their original encoding; episodes extracted from shared RH20T recordings use lossless H.264.

Annotations

Each line in train.jsonl, val.jsonl, or test.jsonl describes one episode. The supplied split membership is fixed.

Field Contents
episode_id Original episode identifier, unique across all splits.
dataset, source_family Source subset and source dataset family.
instruction Original full-episode language instruction.
fps, num_frames Episode frame rate and number of frames.
views Camera IDs, names, relative video paths, FPS, lengths, and primary-view flags.
actions Action IDs, English labels, start/end frames, and hand assignments when annotated.
missions Mission IDs, English labels, start/end frames, and member action IDs.
temporal_schema Serial or overlapping atomic action intervals.

Frame indices are zero-based, and both ends of each interval are inclusive. Video paths are relative to the dataset root. Every video begins at episode frame zero, so each view's offset is zero. A mission's action_ids connect its label to the atomic actions it contains. Four bimanual episodes contain overlapping actions with hand values of left, right, or both.

meta/info.json records split counts, source composition, video properties, and the hashes of the frozen source annotation files. meta/training_run.json identifies the completed runs and their annotation, adapter, and evaluation hashes. meta/checksums.sha256 records the SHA-256 digest of each release file.

Download

Download the dataset together with its license and attribution files:

hf download viamodelanddataset/vi_amo_v2 --repo-type dataset \
  --revision main --local-dir data/vi_amo

Train VIM

Use the current VIM repository and install its environment. From the VIM repository directory, download the base model and prepare the manifests:

hf download unsloth/Qwen3.5-27B \
  --revision 68443620da0c3847c1ba5042b5dcf1e78c42bfba \
  --local-dir checkpoints/qwen3.5-27b

vim-prepare --input data/vi_amo --output data/vi_amo/prepared

vim-train --config configs/vim27b.yaml \
  --model checkpoints/qwen3.5-27b \
  --data data/vi_amo/prepared --output outputs/vim27b

Preparation derives ongoing-action and ongoing-mission targets and the completed actions within each mission, using actions that end before the current action starts. Training reads the videos and samples observations according to the selected VIM configuration. The original full-episode instruction is retained as dataset metadata.

Sources

Source Episodes Actions Missions
Berkeley Fanuc 27 48 32
Bridge V2 3,091 4,950 3,445
CMU Play Fusion 224 725 448
DROID 2,041 6,372 4,452
FastUMI-100K 26 100 52
Fractal (RT-1) 579 656 579
Jaco Play 68 68 68
RH20T cfg 5 95 224 143
HYDRA 3 19 9
TACO Play 403 403 403
UT Austin MUTEX 43 88 71

License

Source materials retain their original licenses. See License and Attribution for licensing, attribution, and modifications.

Downloads last month
121

Models trained or fine-tuned on vi-amo/vi_amo_v2