Robot foundation models compared: GR00T, Gemini Robotics, π, Helix and the rest

The leading robot foundation models, as of September 28, 2026, are NVIDIA’s Isaac GR00T N1.7, Google DeepMind’s Gemini Robotics 2, Physical Intelligence’s π0.7, Figure’s Helix 2.5, Generalist’s GEN-1.5, Skild AI’s S1, AgiBot’s GO-2 and Unitree’s UnifoLM-WLA-1.0. Only GR00T N1.7 ships fully open, commercially usable weights for its current version; the others are closed, partner-only or partly released.

Why it matters: a humanoid maker choosing a model is choosing a supplier. Open weights mean it can fine-tune and ship on its own hardware; closed models mean a partnership, and a dependency, with the model’s owner.

Which robot foundation models are available in 2026?

Eight developers have published a current, named generalist robot model. Most are vision-language-action (VLA) models: a vision-language backbone that outputs motor commands instead of text, as set out in our VLA explainer. The table lists each model’s latest first-party version and how you can get it.

Robot foundation models compared, as of September 28, 2026:

Model	Developer	Latest version	Open weights?	License	Parameters	Robots shown or supported	Access
Isaac GR00T	NVIDIA	N1.7 (early access Apr 17, 2026; now GA)	Yes	Code Apache 2.0; weights NVIDIA Open Model License, commercial use allowed	3B	Unitree G1 (incl. whole-body), Franka Panda, WidowX, Google Robot, DROID; custom embodiments	Hugging Face download; backbone is gated
Gemini Robotics	Google DeepMind	Gemini Robotics 2, ER 2, On-Device 2 (Jul 30, 2026)	No	Proprietary	Not disclosed	Apptronik Apollo 2, Franka Duo; bi-arm robots for On-Device 2	VLA: early-access partners; ER 2: Google AI Studio
π (pi)	Physical Intelligence	π0.7 (Apr 16, 2026)	Partly: π0, π0-FAST, π0.5 only	Apache 2.0 (openpi)	Not disclosed for π0.7	Open checkpoints for ALOHA, DROID, LIBERO; π0.7 shown on UR5e arms	openpi on GitHub; newer models via partner program
Helix	Figure AI	Helix 2.5 (Sep 17, 2026)	No	Proprietary	Not disclosed	Figure humanoids	Internal only
GEN	Generalist	GEN-1.5 (Aug 19, 2026)	No	Proprietary	Not disclosed	Multiple embodiments and end effectors (company-stated)	Early-access partners
Skild Brain	Skild AI	S1 (self-play results Sep 23, 2026)	No	Proprietary	Not disclosed	Unitree G1 (soccer: trained in Isaac Sim, played on real robots)	No public access terms
Genie Operator	AgiBot	GO-2 (Apr 9, 2026)	GO-1 only (Sep 19, 2025)	GO-1: CC BY-NC-SA 4.0 (non-commercial)	Not disclosed	AgiBot robots; GO-1 trained on AgiBot World data	GO-1 on Hugging Face
UnifoLM-WLA	Unitree	1.0 (Sep 10, 2026)	Partly: 4B backbone and ER-Flow; WLA-Base not yet	Apache 2.0	~6B	Unitree G1; grippers and five-finger hands	GitHub and Hugging Face
Optimus model	Tesla	None published	No	–	–	Optimus	None

We list versions only where the developer has published them. NVIDIA has said an early-access GR00T 2 will follow “within months”, per its presentation at Humanoids Summit Seoul on September 22–23, 2026; until it ships, N1.7 is the current version.

Which robot foundation models have open weights?

Four developers release weights, and each on different terms. NVIDIA’s GR00T N1.7 is the most usable: a 3B-parameter model built on a Cosmos-Reason2-2B backbone and pretrained with 20,000 hours of human egocentric video. NVIDIA says it is commercially licensable. Its code is Apache 2.0 and its weights sit under the NVIDIA Open Model License. “Open” still has conditions: every GR00T checkpoint loads the Cosmos-Reason2-2B backbone, which is gated on Hugging Face, so users must request access before the model will run. NVIDIA supports it on its Ampere, Hopper, Lovelace, Blackwell and Jetson platforms, and the repository adds ONNX and TensorRT export for deployment on the robot.

Physical Intelligence open-sourced π0, π0-FAST and π0.5 through its openpi repository under Apache 2.0, with the last model release in September 2025. Its newer π0.6 and π0.7 models are not in the repository; the company’s February 24, 2026 partner post offers access by direct arrangement.

AgiBot open-sourced GO-1 and a smaller GO-1 Air on September 19, 2025, but under CC BY-NC-SA 4.0, which bars commercial use. Its April 2026 successor, GO-2, has no announced open release. Unitree licensed UnifoLM-WLA-1.0 under Apache 2.0 but has released it in stages: backbone weights on September 11 and action-expert training code on September 20. As of September 28, 2026, its GitHub checklist still marks the full UnifoLM-WLA-Base checkpoint as unreleased. We tracked that rollout in our UnifoLM-WLA-1.0 report.

How do the closed models compare?

They cannot be compared directly, because no two developers report results on the same tasks, robots or pass criteria. What each has published:

• Gemini Robotics 2 drives a whole humanoid from one checkpoint. Google DeepMind’s own table shows fine manipulation ranging from 92% (unscrewing a bulb) to 32% (dustpan work) on Apollo 2. See our Gemini Robotics 2 analysis.
• Helix 2.5 reached 56% zero-shot success on household chores in 30 unseen homes, against 9% without Figure’s Index pretraining (company-reported). Details in our Helix 2.5 report.
• GEN-1, per Generalist on April 2, 2026, reaches 99% success on simple tasks, against 64% for GEN-0, and was trained on more than half a million hours of physical interaction data. GEN-1.5, on August 19, 2026, added in-context learning of a new task from as little as 12 seconds of demonstration.
• Skild S1 learned dribbling and defending in simulated soccer after the equivalent of more than 140 years of self-play, according to Skild AI’s September 23, 2026 post.

All four are company-reported. None publishes a parameter count. The useful signal is the shape of each result: Google DeepMind and Figure publish failure rates alongside successes, which makes their numbers easier to trust than headline-only claims.

Which model should a humanoid maker use?

For a company that needs to own its stack today, GR00T N1.7 is the default. It is the only current-version model with open, commercially usable weights and a documented path to whole-body humanoid control, through a Unitree G1 embodiment tag and NVIDIA’s SONIC controller. The price is dependence on NVIDIA’s licence terms and hardware.

For research, π0.5 remains the most widely supported open checkpoint, with fine-tuned versions for ALOHA, DROID and LIBERO. GO-1 is useful for teams working on AgiBot data but cannot ship in a commercial product under its licence.

The strongest closed models, Gemini Robotics 2 and Helix 2.5, are either tied to one maker’s robots or limited to early-access partners. Generalist and Skild are selling cross-embodiment models to partners, a bet investors priced when Generalist reached a $3 billion valuation in August. Tesla, the largest announced humanoid programme, has published no model at all. Developers already building on Google’s older preview should note that the Gemini API documentation marks Gemini Robotics-ER 1.5 Preview as planned for deprecation.

The trend is toward closing, not opening. Physical Intelligence stopped releasing weights after π0.5, AgiBot after GO-1, and Unitree’s main checkpoint is still pending. GR00T 2 will show whether NVIDIA keeps its next generation open.