DavidLXu/Push-Any-Letter

Push Any Letter — Dataset v3

124,800 replay-validated successful demonstrations: 52 letters × 24 center-goal orientations × 100 trajectories.

V3 retains all 62,400 v2 episodes and adds 62,400 new episodes at 15°, 45°, …, 345°. The full goal-angle grid now has 15° spacing. Physics, font geometry, state/image/action schema, planner settings and success threshold are unchanged. All newly accepted trajectories are independently action-replayed and checked for duplicates. Added generation uses seed base 2,000,000,000.

Project page · Code · V2 · V1

Files and provenance

V2 and added-angle episodes are merged into 52 per-letter shards in data/. Each shard contains 2,400 trajectories across all 24 orientations. Episode filenames and NPZ contents are unchanged. The original V3 packaging remains available at v3-original; the v1 and v2 releases are unchanged. episodes.jsonl describes all 124,800 episodes and their SHA-256 hashes; SHA256SUMS lists all shard and manifest hashes. statistics.json records measured counts, transitions, coverage, replay errors and the immutable parent-v2 revision. Inherited episodes retain their original release metadata.

Use the repository downloader with --revision v3 for this release; use --revision v1 or v2 to reproduce previous experiments. The existing downloader follows the manifest's shard paths. The expanded archives and extracted dataset require correspondingly more disk space.

Evaluation scope

The inserted angles are now training data, so they cannot also serve as unseen-angle tests for a model trained on the full v3 set. Define fresh held-out angles or explicit train/test angle splits before reporting generalization. V1/V2 policy and world-model experiments remain tied to their original data versions; this release makes no claim of improved learned-model performance. Goals remain centered, with one font and one physics profile.

Schema and attribution

State: [agent_x, agent_y, object_x, object_y, object_angle, goal_x, goal_y, goal_angle]; action: absolute pusher target in pixels. RGB is pre-action uint8 96×96×3. Control is 20 Hz, physics 200 Hz. Positions use a top-left origin and downward-positive y; angles are radians, clockwise on screen. Block friction 1.4, pusher friction 0.6, damping 0.12.

Data is success-filtered planner demonstration data, not optimal trajectories or a representative failure distribution. Exact physics replay requires the matching Arial Bold font; font binaries are not distributed. Code is MIT, data CC BY 4.0; third-party font rights are separate. Created by Lixin Xu (许立昕); cite CITATION.bib and specify version v3. See the code repository for Push-T / Diffusion Policy attribution.

Measured transitions: 94,365,251. Minimum final coverage: 0.950000. Maximum replay error: 0. Parent v2 revision: 422f2eaadc6c76bf6451706da1d046b50ea1e9d8.

V1 policy evaluation report

The completed bilingual technical report covers four state policies and four image policies trained on V1: ten demonstrations per condition. It includes all loss histories, separate 52-case training/test success curves, the full 520-case unseen-angle tests, and replayed examples. State ACT reaches 319/520 (61.35%); image ACT and Flow reach 227/520 and 226/520, respectively, a one-case difference from one training seed. These are V1 results and do not evaluate the larger V2 or V3 datasets. Pin v1 or revision 8c02168906f991bc07ee3648bee6290babf13e47 to reproduce the report.

Related reports

MIT · Diffusion Policy · Flow Matching · ACT