✓ Source card reviewedMITVersion: LeRobot v2
PushT — Diffusion Policy Benchmark
The canonical visual PushT benchmark used by Diffusion Policy, converted to LeRobot with 206 episodes, 25,650 frames, image observations, and planar actions at 10 Hz.
CCea
Published by Cheng Chi et al.LeRobot / Columbia University · Source updated 08 Jun 2026 · Reviewed 17 Jul 2026View dataset files ↗31.6 MB · LeRobot v2Downloads and licence terms are handled by the original publisher on Hugging Face. START HERE
This dataset, explained simply
A simple 2D agent pushes a T-shaped object into a marked target area.
WHAT CAN I TRAIN?✓ Your first diffusion policy✓ Visual imitation learning✓ Action prediction
WHAT DO I NEED?• Python basics• Only 31.6 MB storage• No physical robot required
SUITABILITY✓ Contains robot actions✓ No physical robot required• Check source for language labels
New to the terminology?Episode = one complete attemptObservation = what the robot sees and sensesAction = the command sent to the robot
VERIFICATION RECORD
What EmbodiedData checked
Source statusAvailable at publisherLast checked17 July 2026Dataset versionLeRobot v2Training instructionsBuild-checked; requires user hardware test
Verification confirms the linked metadata at the stated date. It is not a guarantee that upstream files, dependencies or licences will remain unchanged.
READY TO USE IT?
Train your first model with PushT
Follow a beginner-focused path from inspecting one sample to launching a Diffusion Policy training run.
Dataset previewIllustrative image—not an original dataset frame
Preview onlyEXAMPLE OF THE TASKPush the T-shaped block into the marked target region.
DATASET OVERVIEW
What the publisher reports
The canonical visual PushT benchmark used by Diffusion Policy, converted to LeRobot with 206 episodes, 25,650 frames, image observations, and planar actions at 10 Hz.
206Reported episodes
10 HzReported frequency
42m 45sReported duration
LeRobot v2Dataset format
KNOWN CONFIGURATION
Robot, observations and reuse
Robot or capture systemPlanar pushing environment
Reported observationsRGB image, Agent position, 2D actions
ActionsRobot actions are included
EnvironmentSimulation · policy benchmark
Licence shown on source cardMIT
TRUST & LIMITATIONS
What EmbodiedData has—and has not—checked
Original Hugging Face source linkChecked
Summary against publisher cardReviewed
File integrity and missing valuesNot independently tested
Privacy, safety and legal complianceNot independently tested
Always read the original dataset card, licence, limitations and file documentation before training or redistribution.
SOURCE & CITATION
Use the publisher’s citation
EmbodiedData does not invent a citation or dataset version. Open the original card and copy the citation requested by the publisher.
Open original dataset card ↗