Humanoid targets 99.9 percent manipulation with RL

Humanoid says real world reinforcement learning can hit 99.9 percent manipulation reliability. Testing indicates that its KinetIQ Ascend method lets robots move from basic behaviors to deployment ready skills at human speed, skipping months of manual tuning in many cases.

Humanoid, founded in 2024 by Artem Sokolov, has grown to more than 250 engineers, researchers, and innovators across its offices in London, Boston, Vancouver, and San Diego. The company reports it is building commercially viable, scalable, and safe systems for real world applications, with a clear aim to become the No. 1 general purpose industrial humanoid robotics company within two years. In May, Humanoid joined forces with Bosch and Schaeffler to scale the production of its HMND robots, signaling intent to move beyond prototypes toward real manufacturing lines.

KinetIQ Ascend sits at the heart of that push. Humanoid describes Ascend as its real world reinforcement learning method, built on the underlying KinetIQ platform and designed for trial and error learning on actual industrial tasks. The four layer AI framework, the company says, is tailored for deployments rather than offline benchmarks, letting a robot start with a basic behavior and refine it through interaction with real tasks until it is ready for sustained operation. In the company’s framing, Ascend is meant to shorten the loop from concept to production by letting the system continually improve in the field rather than waiting for perfect data sets.

The claim of rapid real world learning dovetails with a broader aspiration: scale. Humanoid’s leadership argues that the race in humanoid automation is largely a question of scale, and real world RL can be a core part of the answer. Jarad Cannon, the company’s chief technology officer, contends that robots that once required months of tuning are now outperforming human demonstrations within days. Documentation indicates Ascend builds on a history of trial and error learning to push skills into real world deployment more quickly than traditional scripting or offline training alone.

What makes this development notable for engineers and operators is less the headline percentage and more the implied shift in practice. The HMND platform is positioned as a test bed for a capability factory, an approach in which new dexterous tasks are incrementally added, tested, and refined in real environments with safety and reliability baked in from the start. This matters because industrial manipulation hinges on reliability under varied conditions, not just peak performance in a lab. The company’s emphasis on real world deployment, safety considerations, and scalable production signals a move from clever demos to systems meant to endure in busy factories.

From a practitioner’s perspective, the reported 99.9 percent manipulation reliability will be judged against real world constraints. First, even with high reliability on a defined set of tasks, long tail tasks and unexpected disturbances often reveal failure modes that require human oversight or rapid re tuning. Second, the transition from a productivity boosting prototype to a production tool typically introduces additional requirements for calibration, maintenance, and change control across a factory’s line. Third, the integrated supply chain and partner network, Bosch and Schaeffler in this case, will be critical to sustaining volume, not just proving capabilities on a few units. Finally, compute and data governance costs, as well as safety and compliance in industrial settings, will shape the ongoing economics of real world RL in humanoids.

If Ascend delivers on its promise, the next milestones will be measurements from pilots outside controlled shops: durability over weeks of operation, performance on a broader set of tasks, and the ability to learn safely from user demonstrations in live lines. The industry will watch whether the capability factory model can scale consistent, safe improvement across multiple sites and task families, or if the gains remain restricted to carefully engineered demonstrations.