Humanoid's reinforcement learning lifts robot task speeds by 42–100%, cuts failures twentyfold

Humanoid built KinetIQ Ascend on top of its previously announced KinetIQ platform by adding trial-and-error learning—a reinforcement learning approach—that lets robots refine their skills directly on industrial tasks. The results are striking: throughput gains ranging from 42 percent to more than 100 percent, coupled with success rate improvements that in one case represent a roughly twentyfold reduction in failures. Across three real-world scenarios—high-speed single-arm picking, cluttered picking with handoff, and complex two-armed handling—the pattern held, suggesting the method generalizes across diverse manipulation challenges.

The company's framing of this as a "capability factory" reflects a shift in how robot capabilities are built. Rather than spending months collecting training data and tuning every new skill by hand, engineers can now start with a basic behavior and let the system iteratively refine it into production readiness. This approach mirrors how large language models scale: Humanoid observed that robot performance improves predictably as training time increases, and simulation experiments support the claim that the method scales all the way to 100 percent reliability. Two additional findings—that improving only the hardest part of a workflow can improve the entire task, and that robots generalize to unseen objects—point to a system robust enough for real deployment, not just lab validation.

No comments yet. Be the first to share your thoughts!