Humanoid says real world RL reaches 99.9% dexterity

Real world reinforcement learning delivers 99.9 percent manipulation reliability, Humanoid says. The London-based startup unveiled KinetIQ Ascend as its latest RL approach designed to push robotic dexterity toward human speed and beyond. Humanoid frames Ascend as a step beyond lab tuning, built to close the gap between trial and deployment on real industrial tasks. The company describes KinetIQ Ascend as a refinement layer atop its four‑layer KinetIQ framework, engineered for real‑world deployment with trial‑and‑error learning that improves directly on actual tasks rather than relying solely on scripted demonstrations.

The key practical claim is that engineers can begin with a basic robot behavior and let reinforcement learning refine it into a deployment‑ready capability, reducing the months of data collection and manual tuning that once defined new skills. Jarad Cannon, Humanoid’s chief technology officer, framed Ascend as a new pathway for scale. “The humanoid race is becoming a question of scale, and real‑world RL can be a core part of the answer,” he said, adding that robots that once required extensive tuning are now outperforming human demonstrations within days in some tasks. The emphasis is not on a single demonstration but on building a capable, repeatable process for real devices in factories.

Humanoid’s broader ambition sits behind Ascend: to become the No. 1 general‑purpose industrial humanoid robotics company within two years. The team has grown fast since Artem Sokolov founded the company in 2024, now counting more than 250 engineers, researchers, and innovators across offices in London, Boston, Vancouver, and San Diego. The company has been moving toward production through partnerships, most notably in May with Bosch and Schaeffler to scale production of its HMND robots. Those partnerships point to a path from a research milestone to a line of market‑ready humanoids deployed on real production floors.

KinetIQ Ascend is marketed as more than a single technique; Humanoid positions it as part of a “capability factory,” a concept the company uses for repeatable, real‑world skill development. Ascend builds on the earlier KinetIQ platform by adding a real‑world reinforcement loop that pushes capabilities from basic behavior toward robust, deployment‑ready operations. The company says KinetIQ Ascend helps robots improve directly on industrial tasks, rather than requiring lengthy, task‑by‑task retraining. In practical terms, this means a manufacturer could, in theory, push a new manipulation task to its HMND fleet and watch the system converge toward reliable, human‑speed performance with less bespoke tuning each time.

From a practitioner’s vantage point, the claim of 99.9 percent manipulation reliability carries real implications for production risk, cycle time, and maintenance. The four‑layer AI framework points to the need for reliable perception, robust control, safe policy learning, and real‑time execution to keep a line moving. The partnership with Bosch and Schaeffler signals a seriousness about bringing these capabilities to scale, not just in a lab but on factory floors where tooling, gripper wear, and task variability all shape results. If Ascend can sustain its performance as hardware evolves and production demands grow, it could sharpen the tradeoffs between data collection, tuning time, and deployed reliability that have long defined industrial robotics.

Watch next for how real‑world RL handles long‑horizon tasks, safety constraints, and integration with HMND hardware across diverse tasks and environments. Humanoid’s progress hints at a future where a capable baseline policy, refined on the factory floor, becomes the standard route to deploying flexible humanoid automation at scale rather than waiting for perfect demonstrations or exhaustive scripted programming.