HumanoidGPT · Part of The Humanoid Group

Open-Source Robot Models

Not every robot model is locked inside one maker's product. Research groups release model weights, training code and large demonstration datasets that anyone can download. This guide explains what is available, what the licences allow and how teams put them to work.

By Arjun Rao · Updated October 2026

What gets released, and what open means

An open release can include several pieces: the trained weights, the code to run them, the code to train them, and the data or a description of it. Many releases include only some of these. The Open Source Initiative's Open Source AI Definition 1.0 sets a demanding bar: users must be free to use, study, modify and share the system, and must receive the parameters, the complete code used to train and run it, and information about the training data detailed enough to build an equivalent system. By that test, a model with downloadable weights but a restrictive licence or undisclosed data is open-weights, not open source. Check which label a release earns before planning around it.

Generalist policies you can fine-tune

Octo is a well-known example of an open generalist robot policy. Its authors trained a transformer on 800,000 trajectories from the Open X-Embodiment dataset and released it so that it can be fine-tuned to new sensors and action spaces within a few hours on standard consumer GPUs. A model like this gives a lab a starting point that has already seen many robots and tasks, which can save effort compared with training from nothing. The catch is that every robot differs, so expect to collect your own demonstrations, adapt the input and output layers, and test carefully before trusting the result on your own hardware.

Open datasets and what they contain

Datasets matter as much as models, because few teams can afford to record demonstrations at scale. BridgeData V2, from researchers at UC Berkeley, Stanford and partners, contains 60,096 trajectories collected across 24 environments on a publicly available low-cost robot, and supports learning from either goal images or language instructions. Larger pooled collections combine data from many labs and robot types. Before using any dataset, check its licence, which robots and grippers it covers, how episodes were collected and labelled, and whether people or private places appear in the footage. Data recorded on simple arms may still help a humanoid's manipulation, but the gap widens as hands and camera positions diverge.

Licences, safety and support

Open models come with no warranty and no field support, so the work of making them safe on your robot falls to you. Read the licence for each component separately, since weights, code and data are often under different terms and some forbid commercial use. Keep a record of the exact versions you used, because a later release can behave differently. Run fine-tuned models inside the same speed, force and workspace limits you would apply to any untested controller. For humanoids such as the Unitree G1, ask the maker whether its low-level interfaces are documented well enough to connect an open model, and whether doing so affects the warranty.

Sources and further reading

The Open Source AI Definition 1.0 — Open Source Initiative (OSI).
The OSI test for calling an AI system open source: freedoms to use, study, modify and share, plus weights, code and data details.
Octo: An Open-Source Generalist Robot Policy — UC Berkeley, Stanford, Carnegie Mellon University and Google DeepMind (arXiv).
An open transformer policy trained on 800k Open X-Embodiment trajectories and built to be fine-tuned on consumer GPUs.
BridgeData V2: A Dataset for Robot Learning at Scale — UC Berkeley, Stanford, Google DeepMind and CMU (arXiv).
Open dataset of 60,096 trajectories in 24 environments on a low-cost robot, released with pretrained models.

Common questions

Where can I download open robot foundation models?

Most are published through the authors' project pages and code repositories, linked from the research paper. Start from the paper rather than a third-party mirror, so you get the official weights, licence and instructions, and check for newer versions or known issues before you build on one.

Why do research labs release robot models for free?

Openness lets other groups reproduce results, compare methods on shared benchmarks and build on each other's work, which speeds up research for everyone. Pooling data across labs also produces collections no single team could afford, and released models attract collaborators.

Robot Foundation Models — Large models trained on broad data, then adapted to control robots across many tasks.
Vision-Language-Action Models — Models that see a scene, read an instruction and output the robot's next movements.
Data & Simulation — Teleoperation, demonstrations and simulated worlds: where robot training data comes from.
Evaluating Humanoid AI — How to judge a humanoid's intelligence on your own tasks, not in a demo video.
LLMs as Task Planners — How large language models break a spoken request into steps a robot can actually carry out.
World Models for Robots — Learned simulators and video prediction that let robots imagine the result of an action before they act.
On-Device or Cloud AI — Where a robot's AI runs, and how latency, privacy, connectivity and power shape that choice.
Open-Source Robot Models — Open robot foundation models and datasets: what gets released, what open really means and how to use them.
AI Regulation for Robots — How the EU AI Act, NIST's AI Risk Management Framework and India's AI guidelines apply to robots.
Latest updates — Robot models are now tested in unseen homes and judged on safety, yet benchmarks still show wide gaps. These figures and dated updates show what progressed, what failed and how to read the claims.
About us — What we cover, how we source every figure, and who publishes this site.

Visit The Humanoid Group

People also search for open robot foundation models, open-weights robot policy, Octo robot policy, BridgeData V2, Open Source AI Definition and robot learning datasets.