DavidLXu/Push-Any-Letter

Push Any Letter

6,240 successful planner demonstrations, 4,736,343 transitions, 52 letter shapes. Created by Lixin Xu (许立昕), 2026.

Code and planning method · Citation

This dataset records a circular pusher moving letter-shaped rigid bodies to a target in a 640 × 640 arena. Goals are centered at (320, 320), with 12 orientations spaced by 30°. Each uppercase/lowercase ASCII letter has ten accepted trajectories per orientation. Initial object poses and collision-free pusher positions vary; the full initial glyph is at least 48 pixels from the inner walls.

Contents

Release: center12-friction14-v1. The 52 TAR shards in data/ contain 120 compressed NPZ episodes each. For example, data/U0054.tar contains U0054/angle000/traj00.npz. episodes.jsonl identifies every shard/member, seed, reset pose, duration, coverage and episode SHA-256. SHA256SUMS verifies shards and manifests; statistics.json records aggregate measurements. NPZ files load with allow_pickle=False.

Array	Shape	Meaning
observations	N × 8	Pre-action agent xy, object xy/angle, goal xy/angle
next_observations	N × 8	State after the corresponding action
actions	N × 2	Absolute pusher target xy, pixels, bounded to [28,612]²
images	N × 96 × 96 × 3	Pre-action RGB, uint8, HWC
rewards	N	Environment coverage-based reward
terminated	N	Reached the success threshold
truncated	N	Time-limit flag
timestamps	N	Seconds since episode start, sampled at 20 Hz
metadata	scalar string	JSON with reset options, physics, font hash, planner trace and replay validation

Coordinates start at the top left; positive y points downward. Angles are radians and clockwise on screen. Physics runs at 200 Hz, with 20 Hz control. Released episodes range from 23 to 2,385 steps, averaging 759.03. NPZ payloads total 3,184,552,013 bytes; the TAR download is slightly larger. Images decompress to about 131 GB across the full dataset.

Generation and validation

A contact-sampling model-predictive planner uses privileged simulation state and forward rollouts. It uses 20 candidate contacts, at most 40 replanning iterations and a 2,400-step budget. The run accepted 6,240 trajectories from 6,269 attempts. The release contains accepted trajectories only. Seeds vary across slots and retries, and duplicate state/action sequences are rejected. These trajectories need not be shortest or optimal.

The glyph collision friction is 1.4; pusher friction is 0.6, with space damping 0.12. Each accepted episode was independently replayed from reset using only its actions. Validation checks finite values, bounded actions, wall clearance, penetration, recorded rewards, termination flags, state agreement and final success. The maximum recorded replay state error in the source environment was 0.0. Coverage at termination averages 0.9592 and is at least 0.95. This describes selected demonstrations, not a learned policy's held-out success rate.

The font is Arial Bold, SHA-256 d72db21f9242aedd6b917d8549ad5921766b24d5f8d0becfda2ff4c620b3c2e0. Font binaries are not redistributed. Exact replay and new evaluations require a matching font supplied locally; array/image consumption does not. Metadata uses a portable font name instead of the generation server's path. Dependency/platform differences may affect numerical replay.

Download

From the code repository:

python -m pip install -e '.[data]'
python examples/download_dataset.py --output datasets/push-any-letter
python examples/replay.py datasets/push-any-letter/episodes/U0054/angle000/traj00.npz --headless --font /path/to/Arial-Bold.ttf

The script verifies shard and episode hashes before use. Pin --revision to a dataset commit for reproducibility. These are trajectory archives, not a tabular datasets.load_dataset package; use huggingface_hub.snapshot_download or the supplied script.

Intended evaluation

All ten demonstrations per letter/orientation are used for imitation learning. The code generates ten fixed unseen target angles, separated from the training grid, and a fresh initial state for every letter/angle: 520 test tasks. Test tasks are generated from the simulator; they are not additional held-out demonstration files in this repository. The final checkpoint is selected by a fixed update count. Periodic train/test curves use 52-case subsets and must be distinguished from the full 520-case test.

The benchmark holds goal position and font fixed. It does not establish transfer to arbitrary target positions, new fonts, other characters or real robots. Success filtering favors solvable initial states and planner behaviors. Diverse starts do not guarantee coverage of every possible strategy. The first complete state/image BC, Diffusion Policy, Flow Matching and ACT comparison is pending at release.

Attribution

Dataset: CC BY 4.0, copyright 2026 Lixin Xu. Credit the author and link to this dataset when reusing it. Code: MIT. Neither license conveys rights to third-party font software.

Please cite both dataset and code/method entries in CITATION.bib. Push-T and Diffusion Policy inspired the task; the contact planner and synthetic demonstrations belong to this project. The training implementations also credit Flow Matching and ACT.