ISTA-9Tasks

A compact release of 1,034 new simulated robot trajectories: 764 constant-instruction source episodes and 270 instruction-switch episodes across nine object–destination tasks. ISTA denotes Instruction-Switching Trajectory Augmentation in this release.

All fixed-instruction source episodes accepted in the bulk collection are included. The switch subset contains 10 episodes for each of 9 final tasks × 3 switch phases. No original LangGap recordings, model weights, failed trials, or development-template duplicates are included.

Release status: complete. All 1,034 HDF5 trajectories have been uploaded and checked against the release file-size manifest. The release is approximately 16.75 GB.

来源 / Attribution

The task selection and upstream data comparison use YC11Hou/langgap_full, inspected at revision 5283f87d62450ab9a2479241845cd41ab698958d. Please acknowledge that upstream dataset when using this task setting. This compact release contains newly collected rollouts rather than redistributing its original recordings.

The environments build on LIBERO and its Franka Panda manipulation setting. Underlying simulator code and assets retain their respective terms; this card does not relicense those assets. No new blanket license is asserted for upstream material.

Tasks and composition

Task ID	Object	Destination	Source	Early	Held	Late	Total
0	Black bowl	Plate	93	10	10	10	123
43	Black bowl	Stove	96	10	10	10	126
44	Black bowl	Cabinet	96	10	10	10	126
49	Ramekin	Plate	96	10	10	10	126
71	Ramekin	Stove	99	10	10	10	129
50	Ramekin	Cabinet	98	10	10	10	128
51	Cookie box	Plate	73	10	10	10	103
73	Cookie box	Stove	39	10	10	10	69
52	Cookie box	Cabinet	74	10	10	10	104
Total			764	90	90	90	1,034

• Source: an agent-authored scripted controller executes a complete task with one constant instruction, under varied initial conditions. These are simulated rollouts, not human teleoperation or learned-policy predictions.
• Early / held / late: branching from a recorded source prefix at an early stage, while holding an object, or late in the task. A new instruction is executed from the branch state, including releasing or replacing the originally held object where required. Phase labels represent collector conditions, not equal-duration thirds.
• Switch rows are attributed to the final task B in an A→B episode. The HDF5 language sequence and switch index preserve both instruction segments.

Sampling and split

The full accepted collection contains 764 sources and 11,767 switches. This release keeps all sources and samples only 270 switches, using seed 42. Within every final-task/phase cell, 8 train, 1 validation, and 1 test episode are selected, prioritizing diverse initial tasks and then scene seeds. Existing split assignments are preserved. New-data scene seeds are grouped across splits, so branches of one scene do not appear across training and evaluation partitions.

The data are not a balanced success-rate benchmark: failures were excluded by collection acceptance checks. Task counts, source success counts, or loss values must not be interpreted as learned-policy success rates. This compact subset is not an independent closed-loop policy evaluation.

Files and fields

• data/{train,validation,test}.parquet: lightweight episode index, visible in the Hub dataset viewer.
• manifest.json: detailed index with relative file paths and original split assignments.
• episodes/<split>/task-<id>/*.hdf5: complete RGB/state/action trajectories, preserving original HDF5 bytes.
• task_inventory.json: task-level counts.
• provenance/: original-index hash, sampling recipe, source references and collector acceptance records.
• file_manifest.json: release file sizes for integrity/completeness checks.
• previews/: nine example source-execution videos.
• load_episode.py: a small reader; downloading the dataset does not execute it automatically.

HDF5 includes 7D actions, 8D robot state, two 256×256 RGB views, per-frame language, task ID, instruction segment, phase, and held-object metadata. Control frequency is 20 Hz. Preview video timing is separate from the control sequence. Read language from the episode rather than treating the final-task label as a constant instruction for switch episodes.

The Parquet instruction column is a final-task description for browsing. It does not replace the per-frame language sequence. The nine object–destination tasks are all represented; historical three-seen/six-unseen labels from another experiment do not apply automatically to training on this release.

Load

import json
from pathlib import Path
import h5py
from huggingface_hub import hf_hub_download

repo = "JiachengliXX/ISTA-9Tasks"
manifest_path = hf_hub_download(repo, "manifest.json", repo_type="dataset")
episodes = json.loads(Path(manifest_path).read_text())["episodes"]
episode = next(e for e in episodes if e["group"] == "held")
path = hf_hub_download(repo, episode["path"], repo_type="dataset")
with h5py.File(path) as f:
    actions = f["actions"][:]
    states = f["observation.state"][:]
    language = f["language"][:]

For full snapshot loading, use snapshot_download(repo_id=repo, repo_type="dataset"). Fit normalization on the training partition for your intended sampler. Action chunks near a switch must respect instruction boundaries; do not silently label post-switch actions with the earlier instruction.

Source previews

0 · 43 · 44 · 49 · 71 · 50 · 51 · 73 · 52

Suggested citation

@misc{ista9tasks2026,
  author = {Li, Jiacheng},
  title = {ISTA-9Tasks: Source and Instruction-Switching Robot Trajectories},
  year = {2026},
  howpublished = {\url{https://huggingface.co/datasets/JiachengliXX/ISTA-9Tasks}}
}

Please also cite the original LangGap dataset and LIBERO work when appropriate.