Quaternion-joint robot arm achieves 1000x better precision with RL

A team of researchers from Chulalongkorn University, Thammasat University, and the University of Glasgow has unveiled a novel robotic arm design that uses quaternion joints to create a cable-driven redundant manipulator. The paper, published on arXiv, describes a 4-segment, 8-joint configuration that achieves a broader workspace than existing designs while requiring fewer motors per degree of freedom — lowering hardware costs. The key innovation lies in the control system: they employed Residual Reinforcement Learning (Residual RL) to handle the complex kinematics and non-linearities inherent in quaternion joints. Compared to the state-of-the-art FABRIK algorithm, Residual RL improved positional and orientational accuracy by three orders of magnitude (1000x better).

Beyond precision, the control implementation is notably simpler, as the team details a complete FABRIK process adapted for learning. This advancement is particularly valuable for industrial applications requiring manipulators to navigate highly obstructed workspaces — such as assembly lines, surgical robots, or nuclear decommissioning. The methodology also provides designers with new tools to optimize both the mechanical configuration and the control system for future cable-driven robots. The paper (arXiv:2606.05236) is open access.