The three companies covered throughout this series don't just differ on hardware and training strategy — they've made completely different bets on who gets to actually use their robot AI: NVIDIA publishes GR00T's full model weights under a permissive open license anyone can download and build on, Google DeepMind open-sources the tooling around Gemini Robotics while keeping the model itself closed, and Tesla shares essentially nothing with anyone outside the company. This closes out the series by asking a different question than raw capability: who actually gets to build on top of this technology, and what does each company's openness strategy reveal about its broader business model?
NVIDIA GR00T: Open Weights as a Platform Strategy
NVIDIA's most recent release, GR00T N1.7 Early Access from April 17, 2026, is a 3-billion-parameter vision-language-action model built on a Cosmos-Reason2-2B backbone, available directly on HuggingFace and GitHub under the permissive Apache 2.0 license — meaning any company or developer can download, modify, and commercially deploy it without licensing fees. Its central technical advance, called EgoScale, pretrains on more than 20,854 hours of human egocentric video, and NVIDIA reports what it describes as the first documented scaling law for robot dexterity: going from 1,000 to 20,000 hours of human egocentric training data more than doubles average task completion rates. Early adopters already building on GR00T include AeiRobot, Foxlink, NEURA Robotics, and Lightwheel. NVIDIA's openness here isn't altruism so much as platform strategy: the company profits from chip sales (Jetson Thor, covered earlier in this series) and simulation tooling (Isaac Sim, Cosmos) regardless of which robotics company ends up using GR00T, so giving the model away for free expands the market for everything else NVIDIA sells.
Google DeepMind Gemini Robotics: Open Tools, Closed Brain
Gemini Robotics, built on Gemini 2.0 with a separate reasoning model that "thinks before acting," takes a genuinely split approach. The surrounding developer toolkit, Safari SDK (formerly called the Gemini Robotics SDK), is a fully open-source Python package supporting model training, evaluation, and deployment workflows, including hardware-specific embodiments like Aloha and Apollo. But the underlying foundation model itself remains closed — multiple 2026 VLA model rankings explicitly classify "Gemini Robotics and other DeepMind successors" as closed systems, accessible through partnership or API rather than downloadable open weights. Boston Dynamics announced a strategic partnership with Google DeepMind at CES 2026 to integrate Gemini Robotics into its Atlas humanoid platform, but that's a negotiated partner integration, not a publicly available open-weight release anyone can build on. This "open tools, closed model" pattern echoes a broader approach Google has taken elsewhere: its Gemini CLI coding tool kept its surrounding code under an Apache license while moving the actual backend service to more restrictive terms in 2026 — a useful, if separate, example of how "open source" can describe only one layer of a product while the actual intelligence underneath stays gated.
Tesla: Fully Closed, Zero External Access
Tesla's approach, consistent with the vertically integrated strategy covered throughout this series — from its proprietary AI5 chip to its in-house neural world simulator — extends fully to software access as well. There's no public SDK, no API, and no path for an outside developer or robotics company to build on Optimus's underlying models at all. Every part of Tesla's stack, from silicon to training data to the deployed model, stays entirely inside the company, tightly coupled to Tesla's own hardware and data loop. That's the clearest possible contrast with NVIDIA's platform-driven openness: Tesla's moat depends on nobody else having access, while NVIDIA's business model benefits from as many companies as possible building on its open model.
The Open-Source Underdogs Punching Above Their Weight
The open-source side of this landscape has produced results too significant to dismiss as a lesser tier. Stanford and Toyota Research Institute's OpenVLA, a 7-billion-parameter Apache 2.0 model trained on 970,000 robot demonstrations from the Open X-Embodiment dataset, outperformed Google DeepMind's far larger 55-billion-parameter RT-2-X by 16.5 percentage points across 29 evaluation tasks — a smaller, fully open model beating a larger closed one with seven times fewer parameters. OpenVLA exceeded 1,000 academic citations within just 12 months of release. Physical Intelligence's Pi0, described as a "partly open" model from a startup that sells only the AI "brain" rather than any robot body, and Hugging Face's LeRobot and SmolVLA project, explicitly aimed at running smaller open VLA models on consumer-grade hardware, round out a genuinely competitive open ecosystem that doesn't simply trail the closed frontier labs.
Does Openness Actually Predict Quality?
It's worth ending on a genuine caveat rather than a simple "open is better" conclusion. ICLR 2026 research specifically found no reliable correlation between a vision-language model backbone's score on standard benchmarks and its downstream performance as a robot policy — a model that tops general visual-question-answering rankings doesn't necessarily produce a better-performing robot once deployed. That means neither a model's openness nor its benchmark reputation is a reliable shortcut for evaluating actual robot capability; task-specific evaluation on real deployment hardware remains necessary regardless of whether the underlying model is open or closed, well-funded or scrappy.
Robot AI Access Models at a Glance
| Company / model | Model weights | Developer tooling | Strategic logic |
|---|---|---|---|
| NVIDIA GR00T | Fully open (Apache 2.0, HuggingFace/GitHub) | Open, broad early-adopter ecosystem | Sell chips and simulation tools regardless of who uses the model |
| Google DeepMind Gemini Robotics | Closed — partnership/API access only | Open (Safari SDK), but connects to a closed backend | Control the core model while enabling ecosystem tooling |
| Tesla Optimus | Fully closed, no external access | None publicly available | Vertical integration moat; proprietary end-to-end advantage |
| OpenVLA (Stanford/TRI) | Fully open (Apache 2.0) | Open, academic-first | Academic and independent developer accessibility |
Frequently Asked Questions
Is NVIDIA's GR00T actually free to use commercially?
Yes. GR00T N1.7 Early Access is released under the Apache 2.0 license on HuggingFace and GitHub, meaning developers and companies can download, modify, and commercially deploy it without licensing fees, unlike Gemini Robotics or Tesla's models.
Can developers actually build on Google's Gemini Robotics model?
Partially. Google DeepMind's Safari SDK, the developer toolkit around Gemini Robotics, is open source, but the underlying foundation model itself remains closed, accessible only through partnerships or API access rather than downloadable open weights.
Does Tesla offer any external access to its Optimus AI models?
No. Tesla provides no public SDK, API, or developer access to its Optimus models, keeping its entire AI stack, from chip to training data to deployed model, proprietary and tightly coupled to its own hardware.
Do open-source robot AI models actually perform worse than closed ones?
Not necessarily. Stanford and TRI's open-source OpenVLA, with 7 billion parameters, outperformed Google DeepMind's much larger 55-billion-parameter closed RT-2-X model by 16.5 percentage points across 29 evaluation tasks, and ICLR 2026 research found no reliable correlation between benchmark scores and real robot task performance.