The headline development in computer vision robotics this September is that whole-body perception and control moved from research demo to shipping product, while the sensors feeding those models got cheaper and more reliable. Google DeepMind's Gemini Robotics 2, NVIDIA's Isaac GR00T N1.7 early access, Figure AI's $1B-plus Series C, and IDS's Nion depth camera entering series production together mark a shift from lab benchmarks toward deployable systems — with grasping from the floor still the clearest unsolved problem. The through-line across all of it is evaluation honesty. Two of this cycle's most important releases are not capabilities at all but measurement tools: a safety benchmark that scores refusal rather than success, and a real-robot manipulation competition designed to catch results that only hold in simulation.
Table of Contents
- Gemini Robotics 2 and the arrival of whole-body control
- What the same benchmark numbers say about the limits
- Two new ways to measure robots that do not measure success
- Who can actually use these models
- The sensor layer caught up too
- Localization is now a hardware story
- Frequently Asked Questions
Gemini Robotics 2 and the arrival of whole-body control
Google DeepMind announced Gemini Robotics 2 on July 30, 2026. The previous generation drove arms and hands. This one adds whole-body control, meaning the model decides how the robot's base, torso, and limbs coordinate as a single system, plus advanced dexterity and multi-robot collaboration. Why that matters is easiest to see in one number.
In DeepMind's own evaluations, the model unscrewed a light bulb with a 92% success rate. Unscrewing requires sustained rotation past the range of a single wrist, so the robot has to reposition its whole body mid-task rather than treat the arm as an isolated actuator. "Whole-body" is not a marketing distinction. A model that only commands an arm treats base position as a fixed precondition set by a separate planner. A model that commands both can trade one against the other — stepping closer instead of overextending, or shifting weight to keep a grasp stable.
What the same benchmark numbers say about the limits
The honest reading of Gemini Robotics 2 comes from the lifting results, not the light bulb. With Apollo and Inspire hands, DeepMind reported lifting success of 76.3% from a shelf, 68.4% from a table, and 45.7% from the floor. That is a clean gradient, and it tracks how hard perception gets as the object drops. A shelf presents objects near sensor height against a structured background.
A table is worse. The floor is the hardest case: the object sits at an oblique viewing angle, often partly occluded by the robot's own body, against a large untextured plane that gives depth estimation little to lock onto. For anyone scoping a deployment, that gradient is the planning input. Tasks staged at waist or shelf height land in the model's strong range today. Bin-picking from the floor, retrieving dropped items, and anything requiring low-clearance approach currently fails more often than it succeeds.
- Strong axis: shelf-height and table-height pick-and-place, coordinated rotation tasks
- Weak axis: floor-level grasping, under 50% success in the vendor's own numbers
- Practical response: raise the presentation height of parts and bins before blaming the model
Two new ways to measure robots that do not measure success
DeepMind also released ASIMOV-Agentic, a safety benchmark. Rather than scoring whether a robot completes a task, it scores whether the robot refuses an unsafe action, recognizes a task it cannot complete safely, and asks for human intervention. As MarkTechPost's coverage of the release notes, that is a separate evaluation axis from task success entirely. The distinction has real operational weight.
A model optimized only for completion has no reward for stopping, which is exactly the failure mode that matters on a factory floor or in a home. A high task-success score tells you nothing about what the system does when the task is a bad idea. On the academic side, the 2026 CVPR Embodied AI Workshop ran ManipArena, a real-robot manipulation competition scoring physical reasoning and generalization across 20 real-world tasks. Per the CVPR 2026 robotics summary, it is the field's answer to vision-language-action results that hold up in simulation and then collapse on hardware. If you are evaluating a vendor claim this year, ask which of these two axes the number came from.
Who can actually use these models
Access has genuinely opened. Gemini Robotics ER 2 — the embodied-reasoning layer that plans multi-step tasks and reads real-time video — is publicly available to developers through the Gemini API and Google AI Studio, according to Google's ER 2 announcement. A third-party robot builder can call it without negotiating a partnership. NVIDIA's Isaac GR00T N1.7 is in early access with commercial licensing, extending generalized humanoid skills including dexterous control to production deployments.
At GTC, Jensen Huang previewed GR00T N2, built on DreamZero research, which NVIDIA says succeeds at new tasks in new environments more than twice as often as leading VLA models. Treat that figure as a vendor claim until someone runs it through a real-robot harness like ManipArena. Capital is following the same curve. Figure AI closed more than $1B of committed capital in Series C at a $39B post-money valuation, led by Parkway Venture Capital with Brookfield, NVIDIA, Intel Capital, and Qualcomm Ventures among others, earmarked for its Helix VLA platform and the BotQ production line.
The sensor layer caught up too
Models get the attention, but September's most immediately usable release for most integrators is a camera. IDS put its Nion time-of-flight camera into series production, delivering 1.2-megapixel depth at 30 fps over a 0.3–7.5 m range, with motion-blur and ambient-light resistance and a synchronized RGB set that fuses color with depth, per RoboticsTomorrow's Vision 2026 report. Each of those specs maps to a known failure. Motion-blur resistance matters because time-of-flight sensors smear depth when either the camera or the target moves during integration.
Ambient-light resistance addresses the classic indoor-demo-to-warehouse collapse. Synchronized RGB-plus-depth removes the calibration and timestamp-alignment work that separate color and depth streams force on every integrator. The 0.3–7.5 m range is worth checking against your application. It covers mobile-robot navigation and most bin-picking, but a 0.3 m floor means very close-in manipulation still needs a second sensor.
Localization is now a hardware story
ugo Inc. announced the semi-humanoid ugo Nova on September 16, 2026, designed and manufactured entirely in Japan. That detail is the signal: the foundation models are increasingly bought from a handful of global suppliers, so the part a national robotics industry builds and controls is the body and the perception hardware around it.
This is what the pilot-to-platform shift looks like in practice. The differentiated, localized cost sits in cameras, actuators, and integration, not in the policy model. Buyers evaluating a humanoid should price those separately and ask explicitly which model runs on it, because "our robot" and "our model" are no longer the same claim.
Frequently Asked Questions
Can I build on these models without a corporate partnership?
Gemini Robotics ER 2 is available to developers through the Gemini API and Google AI Studio. NVIDIA's Isaac GR00T N1.7 requires early access with commercial licensing.
What is a VLA model?
A vision-language-action model takes camera input and a natural-language instruction and outputs robot actions directly, rather than passing through a separate hand-coded motion planner.
Why does floor-level grasping fail so much more often?
The object sits at an oblique angle, is often occluded by the robot's own body, and lies against a large flat surface that gives depth estimation few features to anchor on. DeepMind measured 45.7% success there against 76.3% from a shelf.
You Might Also Like
- What Is New With Tutorials and DIY Robotics in September 2026? Latest company releases and research papers and Key Takeaways
- What Is New With Sensors and Perception Robotics in September 2026? Latest company releases and research papers and Key Takeaways
- What Is New With Robotics Companies in September 2026? Latest company releases and research papers and Key Takeaways
