London-based company Humanoid has introduced KinetIQ Ascend, a method for training humanoid robots through trial and error on actual production tasks. According to the developer, this technology aims to achieve 99.9% success rates for manipulations at human or higher speeds.

KinetIQ Ascend expands the KinetIQ framework that underpins Humanoid's robots. This new approach employs reinforcement learning (RL), allowing the system not only to mimic human actions but also to repeat tasks independently, receiving feedback on success or failure and gradually improving its performance.

Humanoid stated that it is deploying RL not just in simulations but directly on real equipment and in production scenarios around the clock. The company claims this is the first published demonstration of end-to-end RL based on vision for production VLA models trained on a real dual-arm humanoid platform in deployment conditions.

Why Humanoid is Betting on RL

Previously, Humanoid, like many other robot developers, trained its systems through imitation of human demonstrations.

“A model that mimics demonstrations cannot exceed the speed or quality of the demonstrator and does not learn the cost of errors,” reads Humanoid's material.

The company believes that the last percentage points of reliability and the shift to speeds exceeding human capabilities require a different approach, and KinetIQ Ascend aims to bridge this gap through practice on real tasks. Humanoid compares this process to scaling large language models: the longer the training, the higher the success rate.

Humanoid tested KinetIQ Ascend on three production tasks:

  1. The robot was tasked with picking steel bearing rings from a container and placing them on a conveyor. According to the company, after RL training, throughput increased by 42%, reaching 412 rings per hour compared to 291 for the baseline model.
  2. The robot picked an item from a container and handed it to a human. Humanoid reported an 85% increase in throughput, a 35% reduction in average episode duration, and an increase in success rate from 80% to 98%.
  3. The robot was required to lift a container from a table using both hands in any orientation. Humanoid stated that after several days of training, throughput rose from 122 to 279 containers per hour, average episode duration decreased from 22.9 to 12.8 seconds, and success rate increased from 77.6% to 98.9%.
Source: Humanoid.

The company also highlighted two additional findings:

  • Improving the most challenging step of an operation can enhance the overall task outcome;
  • The skill transfers to objects the robot did not encounter during RL training.

Humanoid claims that it measured gains not against the old baseline but through parallel A/B comparisons with the current baseline model. This is crucial for real production environments, where results can vary due to lighting, object positioning, equipment wear, and other factors.

Humanoid Builds an Industrial Chain in Europe

According to The Robot Report, the company employs over 250 engineers, researchers, and specialists, with offices in London, Boston, and Vancouver. In May, Humanoid announced a partnership with Bosch, which will serve as a contract manufacturing partner for the HMND 01 in the European market.

This agreement followed preliminary tests in March, where robots autonomously transported boxes from a conveyor to a cart in Bosch's logistics environment in Bül, handling five sizes of boxes with varying heights, shapes, and weights.

“For Humanoid, this agreement is a critical step in our roadmap, bridging the gap between concept validation and large-scale deployment,” said company founder Artem Sokolov.

On May 13, the developer reported a phased binding agreement with Schaeffler. According to Reuters, the plan includes deploying between 1,000 and 2,000 robots at the partner's global manufacturing sites by 2032.

Competition Intensifies

The announcement of KinetIQ Ascend comes amid a growing race in humanoid robotics. American company Figure raised over $1 billion in a Series C round in 2025, valuing the company at $39 billion.

Apptronik announced in February 2026 that it had secured $520 million from B Capital, Google, Mercedes-Benz, PEAK6, AT&T Ventures, John Deere, and Qatar Investment Authority. This round supplemented an initial Series A of $415 million, bringing the total to over $935 million.

China is also actively supporting the industry. According to Reuters, the Chinese government has allocated at least $20 billion for robotics development since 2024. However, the agency noted that actual sales remain limited: about 12,000 humanoid robots were sold last year, primarily for research purposes rather than mass industrial deployment.

Previously, Nvidia, Unitree, and Sharpa introduced a platform for developing and testing humanoid robot skills, featuring the Unitree H2 Plus chassis, Sharpa Wave tactile five-finger hands, Jetson Thor, and Isaac GR00T software.

In June, Tether invested in German NEURA Robotics, which is developing a "physical AI" platform and plans to mass-produce humanoid robots.