There was no single "natural language processing robotics update" in September 2026 — the month's news was commercial and organizational, not technical. The year's substantive releases in language-driven robot control landed earlier: Figure's Helix 02 in January and Google DeepMind's Gemini Robotics 2 on 30 July 2026.
September's contribution was Sam Altman's 2 September statement that OpenAI will "definitely do a humanoid," which is a plan, not a product. The category in question is the vision-language-action model, or VLA: a system that takes camera images and a spoken or typed instruction and outputs motor commands directly, rather than running a hand-coded program. If you are evaluating this technology now, the practical question is not which model is best but which ones you can actually get — and the answer is narrower than the announcements suggest.
Table of Contents
- What actually shipped in 2026, and when
- Why the 95% label figure matters more than the speed gain
- What you can actually obtain today
- The hardware question, and NVIDIA's reference design
- The safety standard that does not exist yet
- What to watch next
- Frequently Asked Questions
What actually shipped in 2026, and when
Gemini robotics 2 arrived on 30 July as three separate models, not one. Google DeepMind's announcement describes a VLA that converts vision plus language into motor control, Gemini Robotics ER 2 for embodied reasoning and multi-minute multi-step planning, and an On-Device 2 model for edge deployment. The headline change is scope: control extended from upper-body tabletop manipulation to whole humanoid bodies. Figure's work moved on two tracks.
Helix 02 was unveiled on 27 January 2026 as a full-body-autonomy successor to the company's original VLA, and Figure demonstrated it autonomously loading and unloading a real household dishwasher. Separately, Figure's logistics update reported adding short-term visual memory across recent frames and implicit stereo vision. That logistics work produced the most concrete numbers available this year. Figure reports package handling fell from roughly 5.0 seconds to 4.05 seconds each — about 20% faster — while correct shipping-label orientation rose from around 70% to about 95%. Both are the vendor's own measurements on its own task.
Why the 95% label figure matters more than the speed gain
A 20% throughput improvement is useful. A jump from 70% to 95% correct label orientation is a different kind of change, because it moves the task from "needs a human checking every package" toward "needs a human checking exceptions." The cost of the remaining 5% depends entirely on what happens downstream when a label faces the wrong way. This is the general shape of the language-robotics problem in 2026.
The impressive demonstrations are real, and the residual error rates are still high enough that the economics depend on how cheaply you can catch failures. A dishwasher loaded imperfectly costs nothing; a mislabeled pallet can cost a shipment. Treat every number in this space as a vendor benchmark unless told otherwise. NVIDIA's claim that GR00T N2 — a world-action-model architecture from its DreamZero research, due by end of 2026 — succeeds at new tasks in new environments more than twice as often as leading VLA models is NVIDIA's own figure, not an independent evaluation.
What you can actually obtain today
The access split is the single most decision-relevant fact in this roundup. Google's release notes confirm that Gemini Robotics ER 2 is publicly available to developers through the Gemini API and Google AI Studio, plus private preview on the Gemini Enterprise Agent Platform. The VLA and On-Device models remain restricted to early-access partners.
In plain terms: you can buy the thinking layer and not the doing layer. ER 2 can reason about a scene and plan a multi-step sequence, but it will not move your robot. Something else has to close that loop. For teams that need an action policy now, the realistic options are open weights: A common pattern follows directly from this split: use ER 2 for high-level planning, and an open-weight policy for the low-level actions it calls.
- NVIDIA's Isaac GR00T N1.x, published in the Isaac-GR00T repository, which consumes video, proprioceptive state and natural-language instructions and outputs continuous actions.
- Physical Intelligence's π0 and π0-FAST, open-sourced as reported by The Robot Report, with π0-FAST noted for better language-following.
The hardware question, and NVIDIA's reference design
Getting a body to run these models on has been a barrier as real as model access. NVIDIA addressed it with an open humanoid reference design pairing a Unitree H2 Plus body, Sharpa five-fingered hands, Jetson Thor onboard compute and open Isaac GR00T models. Unitree is to make it available in late 2026.
The significance is that it is a published specification rather than a closed platform. A reference design lets a lab replicate a known-good hardware stack instead of guessing which sensors and compute a given model family expects — which is where a great many robot-learning projects stall. It does not remove the integration work. Five-fingered hands, edge inference and whole-body control each carry their own tuning burden, and a late-2026 availability date leaves little room for anyone planning around a 2026 deployment.
The safety standard that does not exist yet
Every humanoid discussed here ships without a published safety standard written for it. ISO 25785-1, the Type C standard covering dynamically stable robots including humanoids, is still an ISO committee draft with no publication date announced.
Type C standards are the machine-specific ones — the tier that tells an integrator what compliance actually requires for that class of machine. The practical consequence for anyone considering a pilot: This is the gap between what the demonstration videos show and what a safety officer can sign off on. It is also why most credible deployments in 2026 are in caged or cleared logistics spaces rather than shared human environments.
- There is no humanoid-specific standard to certify against, so your risk assessment is your own work product.
- Vendor safety claims are not conformity claims to a published humanoid standard, however they are phrased.
- Language-driven control widens the input space in a way fixed programs do not — an instruction can be ambiguous, and the resulting motion is not enumerable in advance.
- Physical separation and supervised operation remain the fallback, because nothing better is codified yet.
What to watch next
Three dated items are worth tracking. NVIDIA has put GR00T N2 at end of 2026; Unitree's reference-design robot is slated for late 2026; and ISO 25785-1 has no date at all, which makes its eventual publication the most informative signal of the three. A published Type C standard would change procurement conversations more than any model release. OpenAI's position is the open question.
Altman's 2 September statement — reported in September 2026 — that the company will "definitely do a humanoid" and pursue other form factors is its first stated move into own-brand robot hardware. No model, timeline, or specification accompanied it. The most useful thing to watch is narrower than any of these: whether Google DeepMind opens the Gemini Robotics 2 VLA beyond early-access partners. That single change would move the whole stack from reasoning-plus-workaround to a purchasable pipeline, and it is the one event that would genuinely justify calling something an update.
Frequently Asked Questions
Can I use Gemini Robotics 2 to control a robot right now?
Not the action layer. Gemini Robotics ER 2 is available through the Gemini API and Google AI Studio for embodied reasoning and planning, but the VLA and On-Device 2 models are restricted to early-access partners.
What is the difference between a VLA and an embodied reasoning model?
A VLA turns camera images and an instruction into motor commands directly. An embodied reasoning model like ER 2 reasons about the scene and plans multi-step sequences, but leaves the actual motion to something else.
Are humanoid robots certified as safe?
Not against a humanoid-specific standard. ISO 25785-1, the Type C standard covering dynamically stable robots, remains an ISO committee draft with no publication date.
