PushT Mobile — state demonstrations, v2
20 successful human demonstrations collected by zaringleb in the browser-based
PushT Mobile simulator.
This is a state-only LeRobot v3 dataset: 2,155 transitions at 10 Hz (215.5 seconds).
No images or videos are included. The earlier zaringleb/pusht_mobile dataset
is unchanged.
State and action
observation.state: float32[agent_x, agent_y, block_x, block_y, block_angle], before the action.action: float32[target_x, target_y], held for ten physics ticks (100 Hz), one control step (10 Hz).next.observation.state: the state after that action, including the terminal state.next.done: true on the last transition of each successful episode.- Standard LeRobot timestamps, frame indices, episode indices and task indices are included.
Coordinates are in the 512 × 512 simulator world. Angles are radians, wrapped
to [-pi, pi). Stationary frames are retained; timestamps measure simulation
time. Use episode boundaries when constructing action sequences. All 20
episodes are in the train split; there is no held-out demonstration split.
Simulator and reproducibility
These demonstrations use product mode in pusht-core, with Rapier2D 0.22,
1.5× geometry and a 95% target-coverage success threshold. They are not recordings
from the original pymunk PushT environment.
raw/ contains the original JSON recordings, including initial poses and seeds,
all recorded states/actions, the final state, simulator settings and simulator
binary SHA-256. meta/recordings.json preserves episode metadata and the mapping
to the exported episode indices; meta/collectors.json records attribution.
meta/simulator.json identifies the source commit and simulator hash.
Every action sequence was replayed from its saved initial state in the Python
simulator: 20/20 reached success, with zero difference from all recorded states.
Exported states, actions, timestamps, terminal flags, initial seeds and raw JSON
backups were checked against the originals. See meta/validation.json.
The small collection is intended for state-policy experiments. Replay correctness does not imply that a policy trained on these demonstrations will generalize.
- Downloads last month
- 96