README.md

metadata

license: mit
task_categories:
  - reinforcement-learning
tags:
  - world-models
  - control
  - trajectories
  - lance
  - stable-worldmodel
configs:
  - config_name: default
    data_files:
      - split: train
        path: viewer/*.parquet

LeWM Reacher

DeepMind Control Reacher trajectories from quentinll/lewm-reacher, repackaged for direct inspection and streaming.

Open in the LeWM Dataset Visualizer

Data

• 10,000 episodes
• 2,010,000 timesteps
• 201 steps per episode
• 224 × 224 RGB observations

Core fields are pixels (RGB), action (float32[2]), and observation (float32[6]). The source diagnostics are retained: finger_pos, qpos, qvel, target_pos, reward, score, success, terminated, truncated, render_time, id, and ep_idx. episode_idx and step_idx are canonical indices.

train.lance is the indexed training table. viewer/ is a Parquet mirror for the Hugging Face dataset viewer and the LeWM visualizer.

from stable_worldmodel.data.formats.lance import LanceDataset

dataset = LanceDataset(
    "hf://datasets/fracapuano/lewm-reacher/train.lance",
    num_steps=16,
)
sample = dataset[0]

Provenance

• Source: quentinll/lewm-reacher
• Source revision: e70a080d0d04c6072123c9ebd343acf7fff28dbf
• Source archive SHA-256: 4ff2385e49712caa89f21b8e0a246e2614b621d3f22cf2d1224d845e879a1cc2
• Lance schema implementation: stable-worldmodel revision 85d325cc77449c3889e52b03ee7e561cb2fc0383
• License: MIT

All episode boundaries and numeric fields were validated against the source after canonical float32 normalization. Across 145 sampled frames, JPEG pixel MAE averaged 0.903/255 with a maximum sample mean of 0.925/255. Another 130 rows verified the Parquet mirror against Lance, including identical JPEG bytes.

The stable-worldmodel paper reports faster local and remote trajectory loading with Lance than HDF5 on PushT.