Open Source RoboticsRobots, software & learning

MimicGen

Software

Source-available tool for generating robot demonstrations from a small human dataset; code is limited to noncommercial research/evaluation, while released datasets use CC-BY-4.0. [1] [7]

Installation & first result →Official websiteGetting started

For
Noncommercial robotics researchers studying demonstration generation and imitation learning [1] [7]
Hardware
Documented manipulation environments are simulated [3] [5]
Source rights
Named licenses are documented by part. Includes noncommercial terms. Review licenses
Unknowns
compute requirements not fully established. Review gaps

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware

Requirements · Research details and dates

On this page

Overview

MimicGen creates additional robot manipulation demonstrations from a small set of human examples. It splits demonstrations into object-centric subtasks and adapts those segments to new simulated scenes. It is useful for studying imitation-learning data generation, provided the task exposes the object poses and controller information the method needs. [3] [4]

Is this for you?

Known limitations
  • Source code and derivatives are restricted to noncommercial research/evaluation under section 3.3.
  • Custom tasks require object-pose access, controller/action conversion and subtask segmentation.
  • Default generation aims for 1000 successes unless reduced.
  • robosuite v1.5+ controller conventions are unsupported; simulator versions affect reproduction.

[2] [3] [4] [7]

What you need

Operating system
  • macOS or Linux as documented; not independently tested

[2]

Languages / runtimes
  • Python

[8]

Dependencies
  • Python >=3.6 with Python 3.8 recommended by the installation guide
  • robomimic required; reference commit d0b37cf214bd24fb590d182edb6384333f67b661
  • robosuite reference commit b9d8d3de5e3dfd1724f4a0e6555246c460407daa for provided examples; v1.5+ unsupported
  • MuJoCo 2.3.2 recommended for robosuite reproduction
  • pygame for optional manual subtask annotation

[2]

Host computer

No complete CPU/RAM/GPU minimum established by the inspected installation guide; rendering and policy-training requirements depend on the chosen workflow.

Editorial assessment from cited documentation; no runtime testing.

[2] · Editorial assessment

Installation & first result

Open the upstream getting-started guide [3] · Project documentation [1]

  • Create the documented isolated Python environment and install the source plus pinned dependencies.
  • Download Square source demonstrations and prepare the HDF5 with MG_Square environment annotations.
  • For a small run set NUM_TRAJ=10 and GUARANTEE=False, or use the documented debug option.
  • Generate Square D1 data; inspect success/failure datasets, statistics and playback before training a policy.

[2] [3]

Installation

Install the source in a Python 3.8 environment following the pinned robomimic/robosuite and MuJoCo instructions.

[2]

Hardware, models & data

Supported hardware
  • Documented manipulation environments are simulated; dataset robot variants do not establish real-hardware deployment support

Editorial assessment from cited documentation; no runtime testing.

[3] [5] · Editorial assessment

Inputs and outputs

Human HDF5 demonstrations plus task annotations and JSON generation configuration produce success/failure HDF5 datasets, statistics, logs and playback videos.

[3]

File formats
  • HDF5
  • JSON
  • MP4

[3]

Evaluation

Inspect important_stats.json and successful/failed trajectories. Data-generation success is distinct from downstream learned-policy performance.

Editorial assessment from cited documentation; no runtime testing.

[3] [5] · Editorial assessment

Project notes

Code and dataset permissions are different

The source-code license limits the work and derivatives to noncommercial research or evaluation, with an exception for NVIDIA and its affiliates. This catalog therefore treats MimicGen as source-available, not as unrestricted open-source software. The released datasets separately carry CC-BY-4.0 according to the project README and official dataset card; that dataset license does not remove the code's use limitation. [7] [1] [9]

Reproduce the small example before scaling

The installation guide documents macOS or Linux and recommends a Python 3.8 environment. robomimic is required. robosuite is needed for the included examples, and the guide supplies the exact reference commits for both. It explicitly says robosuite 1.5 and newer are unsupported because controller conventions changed. For reproducing its robosuite workflow, it recommends MuJoCo 2.3.2 and describes differences encountered with other versions. [2]

The Square quick start downloads human source demonstrations, adds annotations using the MG_Square environment interface, generates a JSON configuration and runs Square D1 data generation. Reduce the run to ten attempts with NUM_TRAJ=10 and GUARANTEE=False, or use the documented debug option. Otherwise the default configuration keeps generating until it has collected 1,000 successful trajectories. [3]

Inspect the successful and failed HDF5 files, important_stats.json, logs and playback videos. A data-generation run and a policy trained on its output are separate evaluations. The released datasets were postprocessed for robosuite 1.4 from earlier environments, and the dataset guide warns that visual and dynamics differences can prevent exact reproduction of paper results. [3] [5]

New tasks need an environment interface

MimicGen cannot adapt arbitrary videos directly into robot demonstrations. The documented route needs a robomimic-compatible dataset and environment wrapper, an interface supplying object poses and action/pose conversion, and subtask boundaries from heuristics or manual annotation. The custom-environment tutorial describes that integration work. [3] [4]

This review pins commit 72bd767c255545f462e7ccfb2731f2e5d4c1d9bb. The latest GitHub release listing is v1.0.0, while the README announces a later v1.0.1 update moving dataset hosting to Hugging Face; package metadata still says 1.0.0. The current commit is reviewed to capture the documented hosting route and is not presented as a published v1.0.1 release. [1] [6] [8]

We have not installed the dependencies, downloaded the demonstration payloads, generated trajectories or trained a policy. Complete compute requirements, current training costs and physical deployment support remain unverified.

What we checked

Source researched only; no installation, training, benchmark or physical hardware test performed. No affiliate links or reused media.

Updates

Upstream last updated
Aug 16, 2025 [6]
Latest release
Jul 9, 2024 [6]
Profile updated
Oct 17, 2026
Sources last checked
Oct 17, 2026
Next review
Nov 17, 2026
Reviewed revision
current commit snapshot; latest GitHub release v1.0.0 / 72bd767c255545f462e7ccfb2731f2e5d4c1d9bb · software repository

Source files and licenses

Source-available research tool: code and derivatives are noncommercial research/evaluation only; released datasets separately use CC-BY-4.0.

LayerAvailability / license
softwareavailable · NVIDIA Source Code License (noncommercial research/evaluation only)
Section 3.3 limits use to noncommercial research/evaluation, with an NVIDIA/affiliate exception. Redistribution and derivative conditions also apply. [7]
datasetsavailable · CC-BY-4.0
Dataset card inspected; HDF5 payloads not downloaded. Separate from the source-code license. [9] [1]
documentationavailable · NVIDIA Source Code License (noncommercial research/evaluation only)
Repository materials only; external components retain their own terms. [7]

This profile covers demonstration generation and released demonstration datasets, not a pretrained model distribution.

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.7.0-draft.5

Change history