London-based startup Humanoid announced KinetIQ Ascend on July 5, 2026 — a real-world reinforcement learning (RL) approach enabling industrial robots to reach 99.9% manipulation reliability at human speed and beyond, without weeks of manual tuning. Results were published in a technical report on the company's website.
Key takeaways
- KinetIQ Ascend reaches 99.9% manipulation reliability after just a few days of real-world RL training
- Bin-picking and hand-off task: throughput up 85%, success rate from 80% to 98%
- Bimanual tote handling: success rate from 78% to 99%, roughly twentyfold reduction in failures
- Robot generalizes to objects not seen during training
- Humanoid employs over 250 engineers and signed a production partnership with Bosch and Schaeffler in May 2026
What is KinetIQ Ascend?
KinetIQ is Humanoid's proprietary four-layer AI framework built for real-world industrial deployment. KinetIQ Ascend is the latest addition — a module that uses reinforcement learning conducted directly in live environments.
The previous model required engineers to collect demonstration data, manually tune behaviors, and repeat the process for every new skill. Ascend replaces this with a trial-and-error loop — the robot starts from a basic behavior and improves it autonomously for a specific industrial task.
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 call building a 'capability factory'.
Jarad Cannon, CTO of Humanoid
Three tasks, three results
Humanoid tested KinetIQ Ascend on three tasks. First: picking steel bearing rings from a bin and placing them on a conveyor. The method increased throughput by 42%, with the robot operating at 1.5x the speed of the original human demonstration.
The second task involved picking items from a cluttered tote and handing them to a person. Throughput was up 85%, and success rate jumped from 80% to 98%.
The third, most demanding task was bimanual tote handling — lifting a tote from a table using both arms. KinetIQ Ascend more than doubled throughput and pushed success rates from 78% to 99%. That translates to a roughly twentyfold reduction in failures, all achieved after only a few days of training.
Scaling like language models
Humanoid highlights an additional finding: performance improves predictably as training time increases — a pattern resembling the scaling laws seen in large language models (LLMs). Simulation experiments support the trend, suggesting 100% reliability could be achievable.
Two other effects were also observed: improving only the hardest part of a workflow lifts the whole pipeline, and the robot generalizes to objects it had not seen during training.
Context: the manipulation race
Manipulation remains one of the key bottlenecks in production-grade humanoids. Competitors like Figure AI, Apptronik, and Boston Dynamics are investing in similar learning mechanisms — but publicly reported reliability figures above 99% are rare.
Humanoid, founded in 2024 by Artem Sokolov, employs over 250 people across London, Boston, Vancouver, and San Diego. In May 2026, the company announced a production partnership with Bosch and Schaeffler to scale manufacturing of its HMND robots — giving KinetIQ Ascend results a commercial, not just research, dimension.
Why this matters
Reinforcement learning in industrial environments is hard. The environment is unpredictable, downtime is expensive, and every failure must be captured and processed. KinetIQ Ascend suggests this cycle can be compressed from months to days — changing the economics of deployment significantly.
If RL scaling laws prove as robust as those for LLMs, companies with the most training hours and the widest variety of industrial tasks will have a compounding advantage. Humanoid is positioning itself in exactly that race — with Bosch and Schaeffler as production partners. Whether 99.9% reliability holds outside narrowly defined production-line scenarios remains to be tested. The technical report covers limited task configurations, and the market will verify these claims at scale only through mass deployment.
What's next?
- Humanoid plans to scale HMND robot production with Bosch and Schaeffler — announced May 2026
- The company has committed to further RL scaling research targeting 100% reliability, documented in the technical report
- Full methodology available at thehumanoid.ai
Sources
- The Robot Report — Humanoid says KinetIQ Ascend reinforcement learning approaches human-level dexterity
- Humanoid (thehumanoid.ai) — KinetIQ Ascend: Toward 100% Reliable Manipulation and Superhuman Speed
