Open-Source Robot Models vs Closed Foundation Stacks

Download a robot policy the way you download a language model. That is roughly where robotics arrived in 2025. Hugging Face's LeRobot ships model weights, datasets, and a training loop that runs on a single consumer GPU. Physical Intelligence released the weights for π0, one of the more capable generalist manipulation policies published so far. NVIDIA put Isaac GR00T N1 on the same footing: open weights, plus an open post-training recipe for humanoids. A graduate student and a well-funded lab can now start from a similar checkpoint.

Then look at the other half of the field. Google DeepMind's Gemini Robotics arrives as an API and a short list of partner integrations, not a weights download. Figure, 1X, and Skild keep their policies inside the robot and the company. Two philosophies now run in parallel on the same hardware. The interesting question is not which one wins a demo this quarter. It is which one compounds.

The real split is not "open versus closed" in the abstract. It is a stack of concrete choices: who controls the model, who can inspect it, who can ship it without asking permission, and, underneath all of that, who owns the data the model is only as good as.

What "open" actually means here

Openness in robot foundation models has at least four layers, and almost no release opens all of them: the weights, the training code, the license, and the data. GR00T N1 and π0 are open-weight and open-code, which lets you fine-tune and self-host. That is a large gift. But the corpora they were pretrained on are only partly disclosed, and the licenses vary in what commercial use they permit.

Contrast that with the datasets that are genuinely open. Open X-Embodiment pooled data from 22 robot embodiments across dozens of labs into one collection, on the order of a million trajectories. DROID added roughly 76,000 teleoperated demonstrations captured in the wild across many scenes. RH20T pushed contact-rich multimodal manipulation. These are open-data, and they are the reason the open-weight models exist at all. The lesson: most "open" robot models are open-weight, not open-data, and the two are not the same freedom.

The open stack: LeRobot as a center of gravity

LeRobot is less a model than a gravity well. It standardizes a dataset format, hosts thousands of community datasets and policies on the Hub, and ships small trainable models such as SmolVLA that a lab can run without a cluster. Pair that with cheap, reproducible arms in the $100-500 range and you get something the closed stack cannot manufacture by decree: a shared substrate where a fine-tune from one group is directly loadable by the next.

The payoff compounds. Every dataset uploaded in the common format makes the next model cheaper to train. Every policy published invites a benchmark, a bug report, a fork. This is the same dynamic that made open language models a research default, transplanted onto real hardware, where reproducibility has historically been miserable.

Open weights let a thousand labs start from the same checkpoint. They do not give those labs the same data. That is where the real divergence begins.

The closed stack: vertical integration as a strategy

The closed players are not hoarding out of habit. Vertical integration is the bet. When one company owns the hardware, the teleoperation fleet, the model, and the deployment, it controls a data flywheel: every hour a robot operates in a customer site feeds the next training run. Gemini Robotics leans on Google's broader multimodal models and keeps the weights behind an interface. Figure builds its own VLA for its own humanoid. The advantage is coherence and speed on a single embodiment, plus a commercial moat and a cleaner story on safety and liability, since nobody can fork a policy you never shipped.

The cost is reach. A closed model improves on the data its owner can collect. An open model improves on the data everyone collects. Over a long enough horizon those are very different growth curves, and it is not obvious which curve wins for which task.

Scoreboard: what each stack actually gives you

Who benefits, and where the real bottleneck is

For a research group or a seed-stage startup, open is close to oxygen. Fine-tuning π0 or GR00T on a modest dataset costs a rounding error next to pretraining a generalist policy from scratch, which few outside the largest labs can afford. For a robot OEM the calculus is genuinely mixed: an open base cuts the model bill, but the durable advantage moves to whoever owns the deployment data, so many will fine-tune an open model on private data and keep that fine-tune closed. For the capitalized incumbent, closed stays rational as long as the internal flywheel outruns the community's. And for the field as a whole, the open datasets are the quiet infrastructure: without OXE, DROID, and the egocentric corpora, neither camp would have a base big enough to matter.

Here is what both stacks agree on, whether or not they say it out loud. The binding constraint is not model architecture and not compute. It is data: real, diverse, physically grounded records of bodies doing tasks. Open weights do not help if every lab fine-tunes on the same few thousand trajectories. A closed flywheel does not help if it only ever sees one factory. Robotics does not have its web-scale corpus yet, and the gap between language models and robot models is mostly the gap between a trillion tokens of text and a few thousand hours of embodied demonstration.

Where this is heading

Expect the two stacks to blur before either wins. The likely shape of the next few years: an open base model, fine-tuned on private demonstration data, deployed under a commercial contract. Open where it lowers cost, closed where it protects margin. In that world the leverage sits with neither the model publisher nor the robot maker, but with whoever can produce embodied data at scale and hand it over clean. Watch the datasets, not the checkpoints.