London-based company Humanoid has officially unveiled KinetIQ Ascend — an innovative approach to training humanoid robots, based on trial and error directly in real-world production environments. According to the developer, this technology can boost manipulation efficiency to 99.9% at speeds comparable to or exceeding human performance.
KinetIQ Ascend is an extension of the existing KinetIQ framework, which underpins all Humanoid robots. The key difference of the new approach is the use of reinforcement learning (RL). Unlike simply copying human actions, the system independently repeats a task, receives feedback on success or failure, and iteratively improves its actions. Humanoid runs RL not in virtual simulation, but directly on real hardware, around the clock, and in production scenarios. To my knowledge, this is the first published demonstration of end-to-end vision-based RL for VLA models (vision-language-action) trained on a real dual-arm humanoid platform in deployment conditions.
Why RL instead of imitation?
Until now, Humanoid, like most competitors, relied on learning through imitation of human demonstrations. However, as the company rightly notes, a model that copies demonstrations fundamentally cannot exceed the speed or quality of the demonstrator and does not learn the cost of errors. Achieving the final percentage points of reliability and transitioning to superhuman speed requires a fundamentally different approach. KinetIQ Ascend aims to close this gap through continuous practice on real tasks. The company compares this process to scaling large language models: the longer the training, the higher the success rate.
Humanoid has already tested KinetIQ Ascend on three production tasks with impressive results:
- Grabbing steel bearing rings: After RL training, throughput increased by 42% — from 291 to 412 rings per hour.
- Handing an object to a human: Throughput increased by 85%, the average episode duration decreased by 35%, and success rate rose from 80% to 98%.
- Two-handed container lifting: After several days of training, throughput increased from 122 to 279 containers per hour, the average episode duration dropped from 22.9 to 12.8 seconds, and success rate rose from 77.6% to 98.9%.
It is important to note that Humanoid conducted parallel A/B comparisons with the current baseline model, which is critical for real production environments where results can vary due to lighting, object positions, and equipment wear. The company also identified two key findings: improving the most challenging step of an operation can enhance the overall task outcome, and the acquired skill transfers to objects the robot did not see during RL training.
Industrial chain and competition
Humanoid is actively building an industrial chain in Europe. The company employs over 250 specialists, with offices in London, Boston, and Vancouver. In May, a partnership with Bosch was announced, with Bosch becoming a contract manufacturing partner for producing the HMND 01 model on the European market. Additionally, Humanoid has signed a phased binding agreement with Schaeffler, providing for the deployment of 1,000 to 2,000 robots at the partner's global production sites by 2032.
The announcement of KinetIQ Ascend comes amid an accelerating race in humanoid robotics. American company Figure has raised over $1 billion, Apptronik over $935 million, and China has allocated at least $20 billion to develop the industry since 2024. Nevertheless, actual sales remain limited: about 12,000 humanoid robots were sold last year, mostly for research purposes.
Expert opinion: KinetIQ Ascend is a significant step forward that addresses a key challenge in modern robotics: transitioning from laboratory demonstrations to reliable industrial operation. The ability to learn from its own mistakes in a real environment, rather than in simulation, allows robots to adapt to unpredictable conditions. If Humanoid succeeds in scaling this approach, we will witness not just an evolution, but a true revolution in manufacturing automation, where robots become not merely tools, but full-fledged, self-learning team members.