UK robotics and artificial intelligence company Humanoid has introduced KinetIQ Ascend, a reinforcement learning system intended to improve the speed and reliability of robotic manipulation in industrial settings. KinetIQ Ascend extends the company’s KinetIQ platform by allowing robots to refine previously learned behaviours through trial-and-error training on specific tasks. Humanoid said the approach is designed to reduce the manual tuning and data collection required when developing new robotic capabilities.
The company tested the system on tasks including picking components from bins, handing objects to people and moving containers using two arms. In a machine-feeding test, a robot picked steel bearing rings from a bin and placed them on a conveyor. Humanoid reported that KinetIQ Ascend increased throughput by 42%, allowing the robot to operate at 1.5 times the speed of the human demonstrations used for its initial training.
In another test, involving the removal of items from a cluttered container and their transfer to a person, throughput increased by 85%, while the reported success rate rose from 80% to 98%. The company also tested the system on a two-arm task in which a robot lifted a container from a table. Humanoid said throughput more than doubled and the success rate increased from 78% to 99% after several days of training.
According to Humanoid, the results indicate that performance improved as additional training resources were applied. Simulation experiments also suggested that reliability followed a predictable scaling pattern, although the company’s target of 99.9% manipulation reliability has not been demonstrated across all tasks described.
The testing also found that improving the most difficult stage of a workflow could raise performance across the wider task. Humanoid said robots were also able to handle some objects that had not appeared in their training data. Chief Technology Officer Jarad Cannon said the company aims to use real-world reinforcement learning to shorten the process of converting basic robot behaviours into capabilities suitable for industrial deployment.
Humanoid has published a technical report describing the training infrastructure, algorithms and experimental results behind KinetIQ Ascend.
