The London-based company Humanoid has introduced KinetIQ Ascend, an approach to training humanoid robots using trial and error techniques directly on real production tasks. Technology uses reinforcement learning (RL), allowing the system not just to copy a person’s actions but to self-improve behavior through signals about success or failure. According to the developer, this is the first published visual-based cross-RL demonstration for production VLA models on a real two-hand platform.

In tests on three tasks, KinetIQ Ascend showed a significant increase in efficiency: capacity increased by 42-85%, and manipulation success rose from 77-80% to 98-99%. The company emphasized that measurements were performed through a parallel A/B comparison with the current base model, which is important for accounting of production environment variables - lighting, site positions, equipment wear. Also, the skill was transferred to objects that the robot did not see during training.

Humanoid is actively scaling up its industrial rollout: in May, the company announced a partnership with Bosch to produce HMND 01 robots in Europe and signed an agreement with Schaeffler to deploy 1,000-2,000 robots by 2032. Amid growing competition from Figure, Apptronik and Chinese developers, Humanoid makes real-world learning the key to achieving 99.9% reliability of manipulation.