AI helps quantum and robots
NVIDIA researchers are using an Ising‑model AI approach to stabilize quantum computers through calibration and error correction. (x.com) Separately, Google published work showing AI improvements in robots’ spatial reasoning and the ability to read instruments in real environments. (x.com)
Quantum computers drift out of tune, and robots still struggle to understand messy physical spaces. Nvidia and Google this week each published new artificial intelligence systems aimed at those two bottlenecks. (nvidia.com) A quantum computer works by steering fragile quantum bits with carefully tuned control settings, and those settings have to be recalibrated as hardware changes. Nvidia said on April 14 that its open Ising model family automates that tuning and also speeds the error-correction step that flags and fixes hardware mistakes. (nvidia.com) Nvidia’s release includes a 35 billion parameter vision-language model for calibration and two smaller decoding models for surface-code error correction, with 0.9 million and 1.8 million parameters. The company said the decoding models deliver 2.5 times the speed and 3 times the accuracy of its benchmark approach. (nvidia.com; github.com) The calibration side starts with a simpler problem: reading the plots that quantum engineers already use to tune devices, the way a mechanic reads gauges on a dashboard. Nvidia researchers introduced QCalEval on April 14 as a 243-sample benchmark spanning 87 scenario types from 22 experimental tasks to test how well models can interpret those plots. (research.nvidia.com) Error correction is the other half of the job. In today’s leading surface-code designs, software has to infer where noise hit the machine from patterns of detector signals, and Nvidia’s decoding paper describes an artificial-intelligence “pre-decoder” that removes many local errors before handing the rest to a conventional decoder. (arxiv.org) Google’s robotics update tackles a different physical-world problem: knowing what a robot is looking at and whether a task is actually done. Google DeepMind said on April 14 that Gemini Robotics-ER 1.6 improves spatial reasoning, multi-view understanding, task planning and success detection, and adds instrument reading for gauges and sight glasses. (deepmind.google) The model is a reasoning layer rather than the part that directly moves motors. Google said Gemini Robotics-ER 1.6 can call tools such as Google Search, vision-language-action models and user-defined functions, and that the instrument-reading work grew out of collaboration with Boston Dynamics. (deepmind.google) Google framed pointing as the building block for the rest of the system, because a robot has to identify exact objects and locations before it can count items, compare sizes or plan a grasp. The company said the new model improves on Gemini Robotics-ER 1.5 and Gemini 3.0 Flash on pointing, counting and success-detection benchmarks, and made Gemini Robotics-ER 1.6 available through the Gemini Application Programming Interface and Google AI Studio. (deepmind.google) Both projects focus less on flashy demos than on the routine reading and interpretation work that keeps machines useful in the real world. For quantum hardware, that means staying calibrated long enough to compute; for robots, it means understanding a room, a gauge or a partly finished task before taking the next step. (nvidia.com; deepmind.google)
