---
license: other
license_name: mixed-per-source
license_link: https://huggingface.co/datasets/diffusion-gif/trick-prediction-simulation-data/blob/main/LICENSES/README.md
language:
- en
tags:
- robotics
- trajectory-prediction
- simulation
- privacy-minimized
- multiple-licenses
task_categories:
- other
configs:
- config_name: trajectories
  default: true
  data_files:
  - split: train
    path: data/train.jsonl
  - split: heldout
    path: data/heldout.jsonl
- config_name: public-rationales
  data_files:
  - split: train
    path: data/public_rationales.jsonl
---
# Public simulation observations, EEF trajectories and teacher rationale drafts

**17,421 prepared simulation records with actual eight-frame observations**:
14,253 train / 3,168 heldout, from RoboTwin2.0 (five robot bodies) and
DexMimicGen (bimanual simulation). This is a usable visual subset of a frozen
research collection, **not its complete 356.85GB original-media export** and
not a mirror of the complete upstream HDF5/asset datasets.

The 6,219 accompanying public teacher rationales include 3,465 machine-screened
weak-supervision drafts and 2,754 held-for-review replies. Human-approved gold
count is zero; faithfulness and robot-policy/paper accuracy are not verified.
No model inference/training or new teacher API calls were performed here.

## Contents and protocol

- `data/train.jsonl`, `data/heldout.jsonl`: anonymous IDs, public task/body/source
  labels, split, coordinate arrays and tar/member references.
- `data/archives/*.tar`: NPZ observations, each with `f0` ... `f7` containing the
  **original prepared JPEG byte arrays**. All frames are observed-history
  frames, not future footage or model predictions. No resizing/re-encoding.
- `data/public_rationales.jsonl`: only published rationale text, anonymous
  sample ID and explicit draft/review status; no raw API/account metadata.
- `MEDIA_MANIFEST.jsonl`, `SHARDS.json`, `PUBLIC_RELEASE_MANIFEST.json` and
  `CHECKSUMS.sha256`: per-member/per-frame/whole-file checksums.
- `tools/`: export, schema validation and a minimal CPU sample reader.

Coordinates are normalized-1000 image-grid XY. `history` and `future_delta`
are arm-first `[forecast_arm_count][8][2]`; `anchor` is the last observed XY.
Absolute future XY = anchor + future_delta. Each forecast arm has exactly one
fixed local EEF origin (`left_eef_origin` or `right_eef_origin`), not five
legacy body points. The source split and prepared projection protocol are
retained and are not asserted to be the official upstream benchmark split.
Source episode-group IDs are pseudonymized and remain separate across splits.

**The published legacy SFT step900 / GRPO step250 model weights were not trained
on this newer joint/fixed-EEF release.** Reproducing those historical weights
requires their original mix6 prompt/multi-point data protocol; do not silently
substitute this data or claim new model results.

## Download and read (no HF login required)

```python
from huggingface_hub import snapshot_download
root = snapshot_download(
    "diffusion-gif/trick-prediction-simulation-data", repo_type="dataset",
    local_dir="./trick-prediction-simulation-data"
)
```

```bash
pip install -r requirements.txt
python tools/validate_release.py --root /your/downloaded/repository
python tools/load_sample.py --root /your/downloaded/repository --split train --index 0
# Optional observed-history GIF; never overwrites an existing output:
python tools/load_sample.py --root /your/downloaded/repository --index 0 --gif sample.gif
```

Use `local_dir` for a physical download folder rather than cache symlinks.
Known local HF download-cache metadata are ignored by the offline validator;
they are not part of the published payload. Dataset Viewer/load_dataset users
can select `trajectories` or `public-rationales` explicitly to avoid mixing the
two different JSON schemas.

## Observed simulation examples (not predictions)

![Observed simulation history](previews/preview_01.gif)

Six observed-history GIFs and their anonymous record references are listed in
`previews/manifest.json`. They show simulation observations, not real people,
real-world camera footage, ground-truth future videos or model predictions.

## Privacy/release boundary

Source record/episode identifiers, original filenames, absolute machine paths,
source timestamps, raw caption/instruction fields, credentials, HMAC key/ID
mapping, account/API response IDs/usage/cost metadata, audio, logs and model
weights are excluded. Every frame was decoded and checked for EXIF and other
non-allowlisted image metadata. NPZ and tar metadata are rebuilt deterministically.

This is a **simulation-only, field-minimized release**. It is not a claim that
all original real-world datasets/videos have been anonymized, nor a guarantee
against source-scene matching for pixels that originate in public simulations.

Ego-Exo4D, FSD and ShareRobot original media remain outside this release pending
separate redistribution permission and full visual/privacy review. Downloadable
source data and model licenses do not automatically grant rehosting rights.

## Licensing and quality

Use the per-record license and `LICENSES/README.md`: RoboTwin records retain MIT
and upstream attribution; DexMimicGen records retain CC-BY-NC-SA-4.0 restrictions.
These are source licenses, not a new universal license grant for all research data.
Checksums/schema validity do not demonstrate rationale correctness, causal
faithfulness, policy safety or reproduced model/paper performance.