UK startup Humanoid launches reinforcement learning system to improve robot manipulation

UK-based robotics and AI company Humanoid has introduced KinetIQ Ascend, the company’s reinforcement learning approach designed to reach 99.9 percent manipulation reliability at human speed and beyond. KinetIQ Ascend builds on the previously announced KinetIQ platform with trial-and-error learning, helping the company’s robots improve directly on industrial tasks. The new system was tested on several tasks, including picking parts from bins, handing objects to humans, and lifting and moving containers using both arms. It has proven effective across a range of manipulation scenarios. In a machine-feeding application where a robot picks steel bearing rings from a bin and places them onto a conveyor, KinetIQ Ascend increased throughput by 42 percent, enabling the robot to operate at 1.5× the speed of the human demonstrations it originally learned from. In a very different task involving picking items from a cluttered tote and handing them to a person, the same approach increased throughput by 85 percent while improving success rates from 80 percent to 98 percent. Across increasingly complex manipulation scenarios, KinetIQ Ascend continued to deliver significant improvements. In a third bimanual tote handling task where the robot lifts a tote from a table using both arms, throughput more than doubled, and success rates rose from 78 percent to 99 percent, representing a roughly twentyfold reduction in failures, with all results achieved after only a few days of training. The results demonstrate that KinetIQ Ascend shows a new way of developing robot capabilities, proving effective across a range of real-world operational tasks, from high-speed single-arm picking to complex bimanual handling. KinetIQ Ascend also demonstrated that robot performance improves predictably as training time increases. It’s similar to how large language models improve as more compute and data become available. The observed scaling trend, supported by simulation experiments, suggests that the company’s method scales all the way to 100 percent reliability. A new approach also revealed two additional findings: improving only the hardest part of a workflow can improve the entire task, and robots were able to generalise to objects they had not seen during training. Jarad Cannon, chief technology officer at Humanoid, said: “The humanoid race is becoming a question of scale, and real-world RL can be a core part of the answer. Robots that once required months of manual tuning are now outperforming human demonstrations within days. “KinetIQ Ascend, our real-world RL method, offers a new approach to developing robot capabilities. Instead of spending months collecting data and manually tuning every new skill, we can start with a basic behavior and allow RL to refine it into a deployment-ready capability – a process we describe as building a ‘capability factory’, which marks how humanoid robots move from impressive demos to tools that industry can actually rely on.” Humanoid outlined all these findings in a new technical report, which covers the full methodology behind KinetIQ Ascend, including the training infrastructure, algorithmic solutions, and a deeper analysis of the results.