PushT — Diffusion Policy Reference Dataset

What is it?

PushT is a simulated 2D manipulation benchmark dataset developed at Columbia University and distributed by Hugging Face as part of the LeRobot library. Released under Apache 2.0, it contains 206 demonstration episodes of a circular end-effector pushing a T-shaped block to a target position in a 2D environment. Despite its small scale, PushT became the canonical benchmark for evaluating diffusion-based robot learning policies following the Diffusion Policy paper (Chi et al., 2023).

Who is it for?

PushT is used by virtually all robot learning researchers as a first algorithm validation benchmark. Before testing a new policy learning method on expensive real-robot datasets, researchers validate on PushT due to its fast evaluation cycle, reproducible environment, and well-established baseline results. It is particularly useful for comparing imitation learning algorithms including behaviour cloning, DDPM-based diffusion policies, and flow matching approaches.

Key specifications

• •Episodes: 206 demonstrations
• •Environment: Simulated 2D (PyBullet / custom gym environment)
• •Task: Push T-shaped block to target position using circular end-effector
• •Format: LeRobot, Parquet
• •License: Apache 2.0 — commercial use permitted
• •Access: Open — Hugging Face (part of LeRobot)

How it compares

PushT is not comparable to real-robot datasets in scale or complexity — it is a benchmark, not a pretraining source. Its value is standardisation: any researcher reporting results on PushT can be directly compared to hundreds of other papers using the same evaluation. It is to robot learning what MNIST is to computer vision — a simple, universal entry-point benchmark.

Limitations and access notes

PushT is simulated and 2D — results do not directly translate to real-robot performance. It is intended as an algorithm validation benchmark, not a pretraining dataset. Apache 2.0 permits unrestricted commercial use.