UK-based robotics and AI company Humanoid has introduced KinetIQ Ascend, its reinforcement learning approach designed to achieve 99.9% operational reliability at human speeds and above.
KinetIQ Ascend is built on the previously announced KinetIQ platform with trial-and-error learning to help the company’s robots improve directly on industrial tasks.
The new system was tested on several tasks, including picking parts out of a bin, handing something to someone, and using both arms to pick up and move containers. This has proven effective across a variety of operational scenarios.
In a machine feeding application where a robot removes steel bearing rings from a bin and places them on a conveyor, KinetIQ Ascend increased throughput by 42 percent, allowing the robot to operate 1.5 times faster than the human demonstration it first learned from.
For the completely different task of retrieving items from a messy tote bag and handing them to someone, the same approach increased throughput by 85 percent and increased success rate from 80 percent to 98 percent.
Across increasingly complex operational scenarios, KinetIQ Ascend continued to deliver significant improvements. A third two-handed tote handling task, in which the robot used both arms to lift the tote off a table, more than doubled the throughput, increasing the success rate from 78 percent to 99 percent and reducing failures by about 20 times. All results were achieved in just a few days of training.
The results demonstrate that KinetIQ Ascend presents a new way to develop robot capabilities and is effective across a variety of real-world operational tasks, from high-speed single-arm picking to complex two-handed operations.
KinetIQ Ascend also demonstrated that the robot’s performance predictably increases as training time increases. This is similar to how large-scale language models improve as more computing and data become available. The observed scaling trends, supported by simulation experiments, suggest that their method can scale to 100% reliability.
The new approach also revealed two additional findings. The overall task can be improved by improving only the most difficult part of the workflow, and the robot can generalize to objects it has never seen during training.
“Humanoid racing is becoming a matter of scale, and real-world RL may be at the heart of the answer,” said Jarad Cannon, chief technology officer at Humanoid. “Robots that once required months of manual tuning are now outperforming human demonstrations within days.”
“KinetIQ, our real-world RL method Ascend offers a new approach to robot capability development. Instead of spending months collecting data and manually tweaking every new skill, you can start with basic behavior and allow RL to refine it into deployable functionality. This is a process we describe as building a “feature factory,” and it shows how humanoid robots move from impressive demonstrations to real-world, industry-trusted tools. ”
Humanoid has outlined all these findings in a new technical report. This report covers the complete methodology behind KinetIQ Ascend, including the training infrastructure, algorithmic solutions, and deeper analysis of the results.
