In context
Human-robot handover tasks require robots to infer cooperative trajectories from human motion, but traditional imitation learning struggles with generalizing to new spatial regions and adapting to trajectory length variations. By 2023, kernelized movement primitives (KMP) offered a promising approach, yet they demanded dense sampling of the entire workspace and lacked flexibility for scale changes, limiting real-time performance and adaptability in dynamic manufacturing environments.
What was reported
Researchers from Southeast University proposed a multi-level kernelized movement primitives (multi-level KMP) algorithm for one-shot imitation learning of human-robot joint trajectories. The method uses three KMP levels: the first learns the nonlinear mapping between human hand and robot end-effector positions in a reference sub-region; the second transfers this mapping to other sub-regions via parameter modulation, avoiding dense sampling of the full hand-position space; the third models and modulates trajectory length to handle scale variations.
Experiments on a UR5 robot demonstrated that, compared with classical KMP, the multi-level approach reduced average inference error from 7.11 cm to 1.85 cm and cut average inference time from 0.138 s to 0.015 s. For joint trajectories with significant scale differences, handover success rate improved from 13.3% to 100%, showcasing enhanced learning and adaptation capabilities.
Why it mattered
This work addressed key limitations in imitation learning for human-robot collaboration, enabling robots to generalize handover skills across the entire interaction workspace with higher accuracy and real-time performance. The ability to handle trajectory length variations and perform one-shot transfer is critical for deploying robots in unstructured manufacturing tasks where human motion is unpredictable, paving the way for more flexible and efficient human-robot teamwork.
"The proposed algorithm improves handover success rate from 13.3% to 100% on the hand-robot joint trajectories with significant scale variance, which outperforms classical kernelized movement primitives algorithm in the learning ability and adaptability."
Source: 《机器人》期刊 (robot.sia.cn) · Published 2023-07-08 · “基于多级核化运动基元的人机交递轨迹模仿学习”