Reuters: UK startup teaches humanoids
- Robotics startup Skild AI showed off “Skild Brain,” a general-purpose AI model meant to run many kinds of robots, including humanoids and factory machines. - The standout detail is the pitch: one shared brain trained in simulation, human-action video, and fleet data instead of task-by-task robot programming. - That matters because humanoids keep stalling on deployment time — and a reusable software layer could make factory rollouts faster.
Humanoid robots are having a very familiar problem. The hardware keeps getting better, but teaching the machines useful work still takes too long. That is the gap Skild AI is trying to close. The Pittsburgh startup, backed by Amazon and SoftBank, showed a new version of its robot software stack this week and said the same “brain” can be adapted across very different machines — from industrial robots to humanoids. (money.usnews.com) ### What is the thing they actually launched? Skild launched what it calls Skild Brain — a foundation model for robotics. Basically, it is the company’s attempt to do for robots what large language models did for chatbots: build one general system first, then adapt it to lots of tasks later. In Reute(money.usnews.com)han one custom setup per machine. (money.usnews.com) ### Why is that a big deal for humanoids? Because humanoids are not failing on flashy demos. They are failing on the boring middle step between demo and deployment. A robot can look great climbing stairs or recovering from a shove, but factory buyers care about something else — how long it takes to m(money.usnews.com)e system. The next robot starts less from scratch. (money.usnews.com) ### How does Skild say it trains this brain? The company says it uses simulated episodes, human-action videos, and then real-world data gathered from robots already running the software. That matters because robotics has a data problem that language AI does not. There is no giant internet-scale archive of robot interactions to scrape. So companies have to fake a lot of the early experience in simulation, then fine-tune on the expensive real-world stuff. (money.usnews.com) ### What did the demos show? The demo clips showed robots climbing stairs, staying upright after being pushed, and picking up objects in cluttered spaces. Those are not random stunts. They test the messy parts of robotics — balance, contact, uncertainty, and adapting when the world does not match the script. Skild also said the model includes power limits so robots do not apply unsafe force around people. (money.usnews.com) ### Is this the same as proving factory performance? Not really. That is the catch. A good video can show capability, but it does not tell you commissioning time, failure rates, uptime, or cost per task. Reuters’ feature gets at the core promise — faster skill transfer into warehouses and factories —(money.usnews.com)is whether a customer can trust it on shift 200. (money.usnews.com) ### Why are investors so interested anyway? Because the upside is huge if the software layer generalizes. Skild raised $300 million in a Series A round that valued it at $1.5 billion, and its backers include Amazon, SoftBank, Jeff Bezos, Lightspeed, Khosla, Menlo, and Sequoia. Investors are betting that the winning robotics company may not be the one with the prettiest humanoid body — it may be the one with the brain that ports everywhere. (money.usnews.com) ### Who would use this first? The near-term customers look industrial, not domestic. Reuters said Skild’s clients include LG CNS and unnamed partners in logistics and other industrial uses. That tracks with the whole sector. Warehouses and factories are structured environments, the labor economics are clearer, and buyers will pay for reliability before consumers ever buy a home humanoid. (money.usnews.com) ### Bottom line? The real news is not that a robot did something cool on camera. It is that Skild is trying to turn robot training into a reusable software problem. If that works, humanoids stop being custom projects and start looking a lot more like deployable products. (money.usnews.com)
