THULab/dino_wm_pusht_noise_2k_lerobot

DINO-WM PushT (noise 2k) (TsFile)

Apache TsFile version of `neiltan/dino-wm-pusht-noise-2k`.

Overview

A PushT (planar block-pushing) demonstration dataset in the LeRobot v2.1 format, of the kind used to train and evaluate DINO-WM-style world models. In the PushT task an agent pushes a T-shaped block to a fixed goal pose; this collection contains 2,000 noisy demonstration episodes of that single task.

Each record is one control frame within an episode: the 2-D agent/end-effector state, the 2-D action applied at that frame, the per-step reward, and episode-termination / success flags.

• —Scale: 2,000 episodes, 239,900 frames total (episode length 69–185 frames, mean ≈ 120).
• —Sampling rate: 30 fps.
• —Tasks: 1 — "Push the T-shaped block to the goal."
• —Observation/action dims: observation.state is 2-D (motor_0, motor_1); action is 2-D (motor_0, motor_1).

Schema (TsFile structure)

• —Time (INT64, milliseconds) — round(timestamp * 1000), restarting at 0 for each episode.
• —episode_index (TAG) — source episode id (0–1999).
• —task_index (TAG) — source task id (always 0 here).
• —frame_index (FIELD, INT64) — frame position within the episode.
• —sample_index (FIELD, INT64) — the source global index column, renamed.
• —next_reward (FIELD, FLOAT) — reward at the next step.
• —next_done (FIELD, BOOLEAN) — episode-done flag.
• —next_success (FIELD, BOOLEAN) — task-success flag.
• —observation_state_0, observation_state_1 (FIELD, FLOAT) — the 2-D state vector, flattened.
• —action_0, action_1 (FIELD, FLOAT) — the 2-D action vector, flattened.

All 2,000 episodes share a single .tsfile; episode_index and task_index are the device (TAG) dimensions. Query one episode with WHERE episode_index=0.

Vector columns are flattened keeping the source name (observation.state → observation_state_0..1, action → action_0..1). The source timestamp column is dropped because it equals Time / 1000 seconds; frame_index is kept.

Usage

Read the .tsfile files with the Apache TsFile Java or Python SDK.

Source & license

• —Original dataset: https://huggingface.co/datasets/neiltan/dino-wm-pusht-noise-2k
• —Author / publisher: neiltan (Hugging Face)
• —Camera videos (observation.image, 96×96 RGB @ 30 fps) are not included in this repository; see the original dataset's videos/ directory for them.
• —License: not declared by the original dataset; please defer to the original.