{
  "id": 2005256,
  "title": "I Saw the Future of AI in a Robot That Can Learn on the Spot",
  "url": "https://urgent.news/2026/08/19/i-saw-the-future-of-ai-in-a-robot-that-can-learn-on-the-spot",
  "topic": "ai",
  "section": "AI",
  "published": "2026-08-19T19:30:00.000Z",
  "source": {
    "name": "Wired Business",
    "slug": "wired-business",
    "url": "https://www.wired.com/story/generalist-ai-robots-learn-like-clever-toddlers/"
  },
  "original_language": "en",
  "account": "I recently had the opportunity to witness the future of artificial intelligence at the offices of Generalist AI, a startup located in Cambridge, Massachusetts. I was in awe of the robot arms they showcased, which demonstrated an uncanny ability to swiftly learn and adapt to various tasks. These robotic arms mastered a range of simple chores including stacking cups and placing blocks in bowls, after merely watching a short instructional video and without any prior specific training for each individual task. One particularly impressive example involved a robot tasked with sweeping a block into a bowl using a dustpan and brush. When the brush was removed, the robot ingeniously repurposed the dustpan as a makeshift brush and used it to flick the block into the bowl. Another demonstration involved a two-armed robot observing a video of someone unzipping a purse and removing some banknotes. The robot then applied this newfound knowledge to unzip a different type of purse and carefully retrieve the currency. What particularly impressed me was the robot's flexibility in changing its grip when it couldn't initially grasp the money. An engineer present remarked, \"Ha, it never did that before,\" highlighting the robot's ability to adapt in real-time. Generalist AI appears to be focused on teaching its robots about the physics of the world, an approach inspired by the innate human sense of physics developed from an early age. This may significantly contribute to the model's ability to transfer learned skills across different scenarios. Some of the company's demonstrations reminded me of how children improvise and experiment when faced with new tasks. The researchers at Generalist have often been astonished by the robots' decisions, such as one instance where a robot chose to sweep up items with a banana when presented with one. This might seem trivial, but physical intelligence remains a significant gap in machine capabilities. The founders and chief scientists of Generalist AI have impressive backgrounds, having previously worked at Google DeepMind and Boston Dynamics on some of the most advanced robotic models. Traditionally, training AI-powered robots to perform various tasks has involved feeding thousands of examples into the model. However, Generalist and other robotics startups are investing heavily in a general robotic model trained by humans. The company builds special gloves equipped with cameras and robot pincers that people use to perform different chores. I observed a crate piled high with several hundred of these grippers destined for workers in Mexico and other locations. While the exact training recipe used by Generalist remains undisclosed, the company has already amassed a substantial amount of high-quality training data. Unlike other companies pursuing smarter robots, Generalist has developed its AI models entirely from scratch, without relying on open-source language models. Danfei Xu, a roboticist at Georgia Tech familiar with Generalist's work, notes that the startup stands out among companies chasing more general robot models. Xu praises their execution and scientific prowess. In addition to gathering a large quantity of high-quality data, Xu believes the company's roboticists are excelling. Xu also suggests that the demonstration thus far indicates Generalist AI is aiming for real-world commercial deployment of its robots. Karen Liu, a roboticist at Stanford University who is familiar with the company, agrees that Generalist's approach to collecting physical interaction data at scale, without tying it too closely to a specific robot, appears promising. Despite these promising signs, Generalist acknowledges that the learning skills of their models are not yet fully reliable. On average, a robot can successfully complete a task it has been shown about 59% of the time, and achieving a success rate of over 99% is the ideal goal. Moreover, it remains unclear how well these skills will generalize to every conceivable task or setting. Nevertheless, the potential for robots to quickly learn new skills in various industries like manufacturing is enormous. During one evening, an engineer demonstrated this potential by stacking cups on a table in front of a two-armed robot, only for the robot to suddenly join in and assist by grabbing and stacking additional cups with its two grippers. The engineer's delight at this unexpected collaboration was palpable.",
  "summary": "During a recent visit to Generalist AI, I watched a robotic arm improvise and use a banana as a tool.",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 2,
    "also_reported_by": [
      {
        "outlet": "Wired",
        "title": "I Saw the Future of AI in a Robot That Can Learn on the Spot",
        "url": "https://urgent.news/2026/08/19/i-saw-the-future-of-ai-in-a-robot-that-can-learn-on-the-spot-2006414",
        "published": "2026-08-19T19:30:00.000Z"
      }
    ]
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}