Computer vision in robotics changed twice this quarter: Google DeepMind's Gemini Robotics 2, released 30 July 2026, pushed vision-language-action models from table-top arms to whole-body humanoid control, and Ultralytics shipped YOLO27 on 13 September 2026, refreshing the open-source detector that most factory and warehouse vision stacks actually run on. A vision-language-action model, or VLA, is a single network that takes camera frames plus a plain-language instruction and outputs robot motion directly, with no hand-written motion planner in between. Why it matters: the frontier moved from "can the robot see it" to "can the robot act on what it sees, with its whole body, safely, near people." What to watch next is less glamorous — the safety standard that would let a buyer certify a walking robot is still unpublished, and the compute bill for training these policies is now measured in billions.
Table of Contents
- What Gemini Robotics 2 actually changed
- Continuous video changes what a robot can catch
- The open-source side — YOLO27 and the unglamorous majority
- What the money says, and what it does not
- The evidence for humanoids in production, and the gap underneath it
- What to do in the next quarter
- Frequently Asked Questions
What Gemini Robotics 2 actually changed
Google DeepMind announced Gemini Robotics 2 on 30 July 2026 as three models rather than one. The first is a VLA that converts camera and language input into whole-body humanoid motion; earlier versions were limited to upper-body and table-top work. The second, Gemini Robotics ER 2, handles multi-step planning. The third is an on-device VLA that can be adapted to a new robot body with a few hours of collected data.
That third model is the one with the shortest path to a shop floor. On-device means the policy runs on the robot rather than round-tripping to a data centre, which removes network latency from the control loop and removes a cloud dependency from a production line. A few hours of adaptation data is a very different procurement story from a multi-week integration. The whole-body claim is not a lab-only result. Apptronik's humanoid publicly demonstrated the whole-body control at the release, which makes it partner-validated on real hardware.
Continuous video changes what a robot can catch
The most consequential detail in the release is easy to skim past. Per the Gemini Robotics ER 2 model page, success and failure detection now runs on continuous raw video rather than on still frames sampled at intervals. Frame-sampled checking catches the aftermath.
Continuous video catches the event — a cup tipping, a part sliding out of a fixture, a grasp slipping — while it is still happening and while a correction is still possible. For anyone who has watched a cell run a full cycle on a part that was misaligned at step two, that difference is the whole ballgame. ER 2 also adds multi-robot coordination and lower-latency orchestration through the Gemini Live API. Coordination at the model level matters most in mixed fleets, where a mobile base, an arm and a humanoid are working the same space and each currently needs its own bespoke integration.
The open-source side — YOLO27 and the unglamorous majority
Most deployed industrial vision is not a humanoid. It is detection and classification on a conveyor, and a large share of it runs on some generation of YOLO. Ultralytics launched YOLO27 at YOLO Vision 2026 on 13 September 2026 at Shenzhen Bay Culture Square, expanding the open-source family across new tasks and architectures under the theme "Open vision, built for the real world." Manufacturing, logistics and robotics were the named deployment targets, which is a signal about where the maintainers expect the pull to come from.
For a team already on an earlier YOLO release, a new generation is an evaluation exercise, not an automatic upgrade — the honest test is your own held-out images from your own line, under your own lighting. The two stories are not in competition. The foundation-model VLAs are answering "what should the robot do," while the detector family is answering "what is in frame, fast, on hardware you already own.".
What the money says, and what it does not
Industrial vision spending is growing on its own steady curve, not on the humanoid news cycle. A3 Market Intelligence reports North American machine vision transactions of $789 million in Q1 2026, up 6.8% year over year — solid growth, not a spike. The humanoid side looks different.
Figure AI said in September 2026 it will spend upwards of $3.5 billion on NVIDIA GPUs within a year. Read that as a diagnosis: the binding constraint in vision-driven robotics has moved from perception accuracy to the compute and training data needed to produce whole-body policies. For a buyer, those two numbers imply different risk profiles. The machine vision line item is a mature market with many suppliers; the whole-body humanoid line item depends on a vendor sustaining a capital program most of its competitors cannot match.
The evidence for humanoids in production, and the gap underneath it
The strongest sustained commercial record belongs to Agility Robotics. Per the AWS case study on Agility, Digit has moved more than 100,000 totes in GXO-operated fulfillment work, and the Digit 5 generation is specified to lift up to 50 pounds and run over 20 hours a day near people. Tote counts from a paying logistics operator are a far better signal than a demo video. The gap is certification.
ISO 25785-1 — the first safety standard covering dynamically stable robots, meaning bipeds, quadrupeds and balancing wheeled machines, including armed versions — is still at committee and working-draft stage, with publication expected in 2026–2027. US project leadership sits with Agility Robotics, Boston Dynamics and A3. Practically, that means a buyer cannot certify a humanoid against ISO 25785-1 today. Treat the draft as emerging state of the art and a useful checklist for supplier questions, not as a conformity basis. If a vendor implies otherwise, ask them to name the published standard and its edition.
What to do in the next quarter
A short list of moves that are actionable now rather than after the standards settle: The next scheduled checkpoint is close. A3 announced on 1 September 2026 that its Advanced Vision & AI Conference runs 23–24 September 2026 at the Santa Clara Marriott, with roughly 25 speakers from Waymo, NVIDIA, Tesla, Intrinsic, Intel, FANUC, Cognex and Siemens. The agenda — sensor calibration, physical drift and industrial inspection — is a useful corrective to the humanoid coverage: drift is what actually breaks a vision cell six months after commissioning.
- Benchmark YOLO27 against your incumbent detector on your own line images before committing to a migration; accept the new generation only if it wins on your data.
- Ask humanoid vendors for deployed-hours and throughput figures from a named customer operation, in the style of the Digit tote count, rather than lab benchmarks.
- Write the safety case yourself against the ISO 25785-1 draft scope, and re-check it when the standard publishes — expected 2026–2027.
- Treat on-device VLA adaptation time as a procurement question: ask how many hours of data on your hardware, collected by whom, and who owns it afterward.
- Budget the compute, not just the robot. If your integrator is fine-tuning whole-body policies, ask where that training runs and who pays for it.
Frequently Asked Questions
Does Gemini Robotics 2 require a cloud connection?
Not for all three models. One of the three is an on-device VLA, which runs on the robot and can be adapted to a new robot body with a few hours of data, per Google DeepMind's 30 July 2026 announcement.
Can I certify a humanoid to a robot safety standard today?
Not to ISO 25785-1. It is the first standard aimed at dynamically stable robots and remains at committee/working-draft stage, with publication expected 2026–2027.
Is industrial machine vision growing as fast as the humanoid headlines suggest?
No. A3 recorded $789 million in North American machine vision transactions in Q1 2026, up 6.8% year over year — steady growth on an established base.
You Might Also Like
- Simulation Digital Twins Robotics September 2026 Update: What Changed, Why It Matters, and What to Watch Next
- Sensors and Perception Robotics September 2026 Update: What Changed, Why It Matters, and What to Watch Next
- Robotics Companies September 2026 Update: What Changed, Why It Matters, and What to Watch Next
