KinetIQ Ascend Analysis: 99.9% at Human Speed

A robot picking steel bearing rings out of a bin is not glamorous. It is the warehouse equivalent of washing one fork forever, except the fork is slippery, reflective, and financially important. According to RockingRobots, Humanoid tested KinetIQ Ascend on a machine-feeding task where a robot picked steel bearing rings from a bin and placed them on a conveyor. That is exactly the kind of boring task robotics AI has to survive before anyone should let it near the word autonomy without a chaperone. Humanoid’s new claim is bigger than one bin of metal donuts. The company says KinetIQ Ascend is a reinforcement learning approach meant to push robotic manipulation toward 99.9% reliability at human speed and beyond, according to The Robot Report. Translation: the robot should not just copy a human, it should practice, fail, adjust, and eventually stop treating every awkward object pose like a philosophical emergency.

The claim, according to The Robot Report The Robot

Report says London-based Humanoid introduced KinetIQ Ascend as an RL approach designed to reach 99.9% manipulation reliability at human speed and beyond. That number matters because industrial manipulation is brutally unforgiving: 98% can look cute in a demo, then become a confetti cannon of downtime when repeated thousands of times. The company is also not presenting this as a toy lab trick; The Robot Report notes Humanoid is building humanoid robots for general-purpose industrial use and was founded by Artem Sokolov in 2024. Jarad Cannon, Humanoid’s chief technology officer, framed the pitch directly in The Robot Report: “The humanoid race is becoming a question of scale, and real-world RL can be a core part of the answer,” and “Robots that once required months of manual tuning are now outperforming human demonstrations within days.” That is a spicy sentence, and also the right battlefield. Manual tuning is where robot projects go to become artisanal software pickles.

The training loop, according to RockingRobots

RockingRobots reports that KinetIQ Ascend extends Humanoid’s KinetIQ platform by letting robots refine previously learned behaviours through trial-and-error training on specific tasks. That is the key technical move: start with something the robot can already attempt, then let real task feedback shape better behavior. It is reinforcement learning in the messier place where friction, lighting, tolerances, and object weirdness conspire like a unionized chaos department. The task list is useful because it avoids the usual robotics demo trap, where a robot performs one clean motion once and the internet declares victory before the gripper has cooled down. RockingRobots says Humanoid tested the system on picking components from bins, handing objects to people, and moving containers using two arms. Those are not exotic magic tricks, but they are exactly the kind of repeatable manipulation chores that decide whether a robot earns floor space or becomes a very expensive hat rack.

Why real-world RL is the interesting bit, according to The Robot Report The

Robot Report’s coverage makes clear that Humanoid is positioning real-world RL as a scaling answer to the manual tuning bottleneck. In robotics, that bottleneck is not just annoying, it is structural. Every new fixture, bin, part shape, lighting condition, or human handoff can demand careful adjustment, and suddenly your general-purpose robot needs the emotional support spreadsheet of a custom automation project. The AI angle is that imitation alone gives you a ceiling: the system learns to reproduce behavior, but not necessarily to discover faster or more reliable strategies under actual consequences. Humanoid’s pitch, as summarized by The Robot Report and RockingRobots, is that trial and error on production-style tasks can close the last reliability gap. That does not mean you toss a humanoid into a factory and let it learn like a caffeinated toddler. It means the learning system has to be constrained, measurable, and boringly safe, which is where the real engineering lives.

What to watch next, according to RockingRobots and The Robot Report

RockingRobots describes KinetIQ Ascend as intended to reduce manual tuning and data collection when developing new robotic capabilities. The Robot Report adds the headline ambition: 99.9% manipulation reliability at human speed and beyond. The useful question now is not whether the demo looks impressive, but whether reliability holds across longer runs, messier part distributions, and task changes that were not lovingly prepared for camera day. For builders, the takeaway is practical: real-world RL is becoming a serious knob in the robot training stack, not just a paper benchmark wearing safety goggles. If Humanoid can show repeatable results beyond selected tasks, KinetIQ Ascend becomes a signal that industrial humanoids may be trained less like hand-coded automatons and more like systems that improve through structured practice. Until then, keep the optimism, ask for the failure curves, and remember that 99.9% still has to earn every nine.