London-based company Humanoid has introduced KinetIQ Ascend — an innovative approach to training humanoid robots, based on trial and error directly in real-world production environments. The developer claims the technology can achieve manipulation accuracy of up to 99.9% at speeds comparable to or exceeding human performance.
How KinetIQ Ascend works
KinetIQ Ascend expands the capabilities of the KinetIQ framework underlying Humanoid's robots. The key difference is the use of reinforcement learning (RL). Unlike passive imitation of human actions, the system independently repeats a task, receives feedback on success or failure, and improves its behavior algorithms based on this feedback.
Importantly, Humanoid runs RL not only in simulation but around the clock on real equipment and in production scenarios. To my knowledge, this is the first published demonstration of end-to-end vision-based RL for vision-language-action (VLA) models trained on a real dual-arm humanoid platform in deployment conditions. This fundamentally changes the approach to skill acquisition.
Why the focus on RL is justified
Previously, Humanoid, like most competitors, trained robots through imitation of human demonstrations. However, this method has a fundamental limitation: a model that copies demonstrations cannot surpass the demonstrator in speed or quality and does not learn from mistakes. According to the company, the final percentages of reliability and the transition to superhuman speed require a different approach. KinetIQ Ascend bridges this gap through continuous practice on real tasks, comparing the process to scaling large language models — the longer the training, the higher the success rate.
Test results
Humanoid conducted tests on three production tasks. In the first, the robot had to pick up steel bearing rings from a container and place them on a conveyor. After RL training, throughput increased by 42%, to 412 rings per hour compared to 291 for the base model. In the second task — handing an object from a container to a human — throughput increased by 85%, average episode duration decreased by 35%, and success rate rose from 80% to 98%. In the third — dual-arm lifting of a container from a table in arbitrary orientation — after several days of training, throughput increased from 122 to 279 containers per hour, average episode duration decreased from 22.9 to 12.8 seconds, and success rate rose from 77.6% to 98.9%.
The company emphasizes that measurements were conducted through parallel A/B comparison with the current base model, not against an older baseline. This is critically important for real production environments where results can vary due to lighting, object positions, or equipment wear.
Expert opinion
Humanoid's breakthrough is not just another step in robotics but a paradigm shift. The transition from passive imitation to active reinforcement learning on real tasks could be the catalyst that brings humanoid robots out of laboratories and into mass production factories. If the company can scale this approach, we will witness a sharp acceleration of automation in industry.