Robotics model posts 99% reliability
A recent robotics model reportedly achieved 99% reliability on tasks it was not trained on, suggesting stronger generalization for manipulation and control. (x.com) The claim implies fewer task-specific demonstrations may be needed to deploy robots in new settings. (x.com)
Robots usually fail outside the narrow setups they were trained on, but Generalist AI said on April 2 that its new GEN-1 model reached 99% success on several simple physical tasks. (generalistai.com) In robotics, “generalization” means a machine can handle a new object, layout, or interruption without being retrained for each one. Physical Intelligence wrote in April 2025 that this is the main barrier between polished demos and robots that work in homes, hospitals, and stores. (pi.website) Generalist said GEN-1 is a multimodal model that turns camera and sensor input into actions in real time, and that it was trained from scratch on 500,000 hours of real-world data. The company said the model can finish some tasks about three times faster than the prior state of the art. (generalistai.com) The company said GEN-1 hit those results with about one hour of robot-specific adaptation data, rather than a separate large demonstration set for each task. Generalist also said the model showed “recovery in unexpected scenarios,” meaning it could adjust when a task went off script. (generalistai.com) Ars Technica reported that the tasks included folding boxes, packing phones, and servicing robot vacuums, which are repetitive jobs that still require precise hand movements. The same report said earlier systems often looked capable in demos but broke down when conditions changed. (arstechnica.com) Generalist is a San Francisco startup founded in 2024 by Pete Florence, Andy Zeng, and Andy Barry, according to Forbes and company trackers. TechCrunch reported in March 2025 that Zeng had left Google DeepMind and that Nvidia had already invested in the startup. (forbes.com) (techcrunch.com) The claim lands in a field that has been chasing robots that can transfer skills across settings instead of memorizing one bench-top routine. In September 2025, the RDT2 team said its model could deploy zero-shot on unseen robot bodies for simple tasks like picking, placing, pressing, and wiping. (rdt-robotics.github.io) Physical Intelligence said in April 2025 that even strong vision-language-action systems still struggled in entirely new homes and cluttered rooms. Its π0.5 model was presented as progress on new environments, but the company said it was “far from perfect.” (pi.website) Generalist’s 99% figure is the company’s own benchmark claim, and the public materials released so far are a blog post and demo videos rather than a peer-reviewed paper. The next test is whether outside researchers and customers can reproduce those success rates on robots, objects, and work cells the company did not choose itself. (generalistai.com)
