Datasets:

ZaidGhazal
/

world-models-eval

	---
	license: mit
	task_categories:
	- robotics
	tags:
	- lerobot
	- libero
	- robotics
	- robot-learning
	- world-model-evaluation
	- imitation-learning
	- vision-language-action
	- policy-evaluation
	size_categories:
	- 100K<n<1M
	configs:
	- config_name: default
	data_files:
	- split: train
	path: data/chunk-/file-.parquet
	---
	![DreamGrasp banner](assets/dreamgrasp-hf-banner.png)
	# DreamGrasp: Processed LIBERO Manipulation Demonstrations

	Does a robot policy's evaluation still mean something if it never touched a real simulator, only a world model's imagination of one?

	This dataset is the shared training data behind that question, a single, ready-to-train release built from LIBERO's manipulation demonstrations (`libero_spatial`, `libero_object`, `libero_goal`). It provides:

	- Fixed, versioned train / validation / test / held-out splits, so every result trained on this data is directly reproducible and comparable across runs
	- Pre-computed action and proprioceptive normalization statistics, shared by every model in the project rather than recomputed per-run
	- Synchronized agentview and wrist-camera video for every episode, converted to LeRobotDataset v3 and ready to drop into LeRobot's data loaders
	- The exact data used to train both the policy and the five-tier world-model family in the DreamGrasp calibration study, start here instead of re-deriving splits and stats from scratch

	Project links: [GitHub repository](https://github.com/ZaidGhazal/world-models-eval)

	## Contents

	- 1,500 episodes
	- 200,485 frames
	- 30 tasks across 3 LIBERO suites
	- 20 FPS
	- 128 × 128 RGB `agentview` and wrist videos (native LIBERO resolution, not upscaled)
	- Panda proprioceptive state, 8 dimensions: end-effector position, axis-angle orientation, gripper state
	- Normalized 7D delta end-effector action plus gripper command

	## Quickstart

	```python
	from lerobot.datasets.lerobot_dataset import LeRobotDataset

	dataset = LeRobotDataset("ZaidGhazal/world-models-eval")

	episode = dataset[0]
	print(episode["observation.images.agentview"].shape) # (T, 3, 128, 128)
	print(episode["action"].shape) # (T, 7)
	```

	Normalization statistics and the frozen split assignment used throughout the DreamGrasp project are versioned in the repo under `configs/norm_stats.json` and `configs/splits.json` — load these rather than recomputing your own if you want directly comparable results.

	## Features

	\| Field \| Description \|
	\| ------------------------------ \| ------------------------------------------------------------------------ \|
	\| `observation.images.agentview` \| 128 × 128 RGB video \|
	\| `observation.images.wrist` \| 128 × 128 RGB video \|
	\| `observation.state` \| float32 `(8,)` proprioceptive state \|
	\| `action` \| float32 `(7,)` action, normalized to `[-1, 1]` using train episodes only \|
	\| `task_index` \| Integer task id, resolved via `meta/tasks.parquet` \|


	## What you can build with DreamGrasp

	- Train and evaluate manipulation policies on a clean, ready-to-use LIBERO benchmark, with splits and normalization already handled
	- Develop and test world models for robotics, using real demonstration data and a fixed held-out set for fair comparisons
	- Research how well simulated or imagined rollouts predict real policy performance, the question behind DreamGrasp and related work
	- Study generalization and distribution shift in imitation learning, using tasks intentionally withheld from training

	## Splits and Normalization

	The LeRobot metadata ships everything as a single `train` split; DreamGrasp uses the frozen episode assignment in `configs/splits.json`:

	- 960 train episodes
	- 120 validation episodes
	- 120 test episodes
	- 300 held-out episodes, from tasks excluded entirely from training — used specifically to test evaluation reliability under distribution shift

	Action and proprioceptive normalization statistics are stored in `configs/norm_stats.json`. Use these exact stats if you want results directly comparable to the DreamGrasp project's own policy and world-model training.

	## Limitations

	- Simulation-only. No real-robot trajectories; sim-to-real transfer isn't evaluated. In exchange, every result is fully reproducible without hardware.
	- Single embodiment. Franka Emika Panda only, so results may not transfer to other morphologies — but embodiment never confounds a comparison.
	- Fixed task scope. 30 tasks across LIBERO's spatial, object, and goal suites; narrow by design so everything runs end to end on a single GPU.
	- Train-split normalization stats. Held-out tasks may see out-of-range action values. Intentional: held-out evaluation stays a genuine test of generalization.

	## Source and Citation

	If you use this dataset, please cite it:

	```bibtex
	@misc{ghazal2026dreamgrasp,
	title={DreamGrasp: Processed LIBERO Manipulation Demonstrations for World-Model Evaluation},
	author={Ghazal, Zaid},
	year={2026},
	howpublished={\url{https://huggingface.co/datasets/ZaidGhazal/world-models-eval}}
	}
	```

	This dataset is converted from the original LIBERO demonstration HDF5 files. If you use it, please also cite the source benchmark:

	```bibtex
	@article{liu2023libero,
	title={LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning},
	author={Liu, Bo and Zhu, Yifeng and Gao, Chongkai and Feng, Yihao and Liu, Qiang and Zhu, Yuke and Stone, Peter},
	journal={arXiv preprint arXiv:2306.03310},
	year={2023}
	}
	```

	The processing pipeline that produced this dataset from the raw LIBERO files is open-sourced in the DreamGrasp [GitHub repository](https://github.com/ZaidGhazal/world-models-eval), alongside the policy and world-model training code and the calibration study itself.
	---

	DreamGrasp is created and maintained by [Zaid Ghazal](https://github.com/ZaidGhazal).