For years the open question about AI-driven robots was whether they could do useful work at all. This week brought three public attempts at an answer. Google DeepMind announced Gemini Robotics 2.5 on Tuesday, November 10, and reported a 96 percent success rate in a proof of concept on Mercedes-Benz production lines, where Apptronik's Apollo 2 humanoid picked and packed automotive parts. On Monday, November 9, at the Conference on Robot Learning, Nvidia introduced GR00T 2 and reported 71 percent average success across 50 tasks with one demonstration each, and 95 percent with five. Also on November 9, UC Berkeley News described Ken Goldberg's "agentic robotics" approach, including a parcel-sorting change that Ambi Robotics applied across 30 percent of its US fleet.
The capability claims are getting bigger. The shared yardstick is not.
Start with the strongest result. DeepMind's own post calls the 96 percent a promising step, and says the pilot ran at sites in Hungary and Germany. Ars Technica's Ryan Whitwam put it in context: "A human doing that sort of repetitive task will get it right virtually every time," he wrote, calling 96 percent "still probably not enough" for a company like Mercedes-Benz. DeepMind director Kanishka Rao told Ars that closing the last four percent will be a "steep road." Robotics head Carolina Parada named the tradeoff: "We can be very good at a particular application at the cost of generalization. Or we can just push on the generalization axis and maybe not get to 99.9 in every application." That is the vendor's own framing, and it is a useful one. The 96 percent comes from DeepMind and its partners, and none of the sources cited here reports an independent measurement. The 30-to-100-percent gains over the previous version come from laboratory testing across 200 tasks, per DeepMind. Ars adds that the model still runs on cloud servers, so movement carries latency, and that the demo video is sped up two times. On safety, Ars reports 98.8 percent on ASIMOV-Jailbreak with Gemini 4 Argon in the loop against 73.5 percent without it. It describes the benchmark as 1,500 prompts, while DeepMind's post says "hundreds." Either way, DeepMind built the test and the model it grades.
Nvidia's numbers also come from Nvidia. Its blog post reports 71 percent and 95 percent, says GR00T 2 completed 40 of 50 evaluated tasks at least once without task-specific demonstrations, and says the model ranks among the top three on four benchmarks, including RoboDojo, according to Nvidia. The post gives headline figures without a task-by-task breakdown. Trade coverage such as The Robot Report repeats them as Nvidia's claims. None of this makes GR00T 2 weak. Learning a task from one demonstration at a useful rate would be a real advance. The point is narrower: from these sources alone, a reader cannot audit it.
The Berkeley work shows a different way to report progress. Instead of training on ever more demonstrations, Goldberg and his collaborators, including Nvidia's Spencer Huang, have AI agents write robot programs, rehearse them in Nvidia's Isaac simulation libraries, and refine them from failures before deployment. Berkeley News reports that Ambi Robotics applied the methods to parcel sorting, with tests showing capacity up by more than 15,000 packages per robot per year, now deployed across 30 percent of its US fleet. The article does not describe independent verification. Still, it is a production metric, packages sorted, rather than a benchmark score. Goldberg, who last year warned of a "100,000-year data gap," is described by Berkeley News as getting "goosebumps" from the new models. That enthusiasm is his own, and the piece shows his work is built on Nvidia's tools.
Three habits travel across all of it. Ask whose scoreboard a robot number comes from, and whether it was measured in a lab or on a line. Ask what the last few percent cost, because the distance between a pilot and a production line is where the engineering budget goes. And where a deployment metric exists, read it next to the benchmark score.
Robots clocked in. Bring your own scoreboard.
Sources
- Gemini Robotics 2.5: Accelerating Physical Intelligence - Google DeepMind (Nov 10, 2026): https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/
- DeepMind reveals Gemini Robotics 2.5 and open safety benchmark "ASIMOV-jailbreak" - Ars Technica (Nov 10, 2026): https://arstechnica.com/google/2026/11/deepmind-reveals-gemini-robotics-2-5-and-open-safety-benchmark-asimov-jailbreak/
- Practice Makes Precise: At CoRL, NVIDIA and Partners Reveal How Robots Rehearse Delicate Work in Simulation - NVIDIA Blog (Nov 9, 2026): https://blogs.nvidia.com/blog/robotics-simulation-advancements-corl/
- NVIDIA launches GR00T 2 and Isaac Lab 3.0 at CoRL - The Robot Report (Nov 9, 2026): https://www.therobotreport.com/nvidia-launches-gr00t-2-isaac-lab-3-0-corl/
- Recent advances in AI are creating a 'paradigm shift' for robotics - Berkeley News (Nov 9, 2026): https://news.berkeley.edu/2026/11/09/recent-advances-in-ai-are-creating-a-paradi
Full disclosure: Ganesh uses Instinct (instinct.ai), an AI assistant, to research, write and publish this blog series.
Top comments (0)