基于ProMPs和PI~2的机器人学习方法

基于传统运动基元模型的机器人学习方法存在学习速度慢、学习结果精度低等问题,为此本文提出一种融合贝叶斯估计算法的概率运动基元(ProMPs)表达和模仿学习框架,同时还利用了基于核典型相关分析(KCCA)的改进型路径积分PI~2策略进行轨迹优化。ProMPs结合贝叶斯推断,为机器人实现有别于示范任务的新任务提供了一个可行解搜索起点,而利用附加泛函指标约束的PI~2算法能让机器人获得平滑的过点轨迹。通过UR5机器人实验平台和V-REP仿真软件对本文方法进行过点试验验证,结果表明,所提出的贝叶斯ProMPs-PI~2学习方法能快速而精准地完成机器人从示范任务到陌生任务的泛化学习,实现机器人新技能的获取。

Robot learning method based on ProMPs and PI2

Traditional robot learning methods based on dynamical movement primitives (DMPs) have such problems as slow learning speed and low accuracy, so this paper proposes a framework for expression and imitation learning by using Bayesian estimation-based probabilistic movement primitives (ProMPs), and applies the policy improvement with path integrals (PI²) strategy improved by kernel canonical correlation analysis (KCCA) to optimize the trajectories. ProMPs in combination with Bayesian inferences provides a starting point searching for feasible solutions of new tasks which are different from the demonstration, and PI² algorithm with additional functional index constraints allows the robot to get smooth via-point trajectories. Via-point tests are carried out on UR5 robot experiment platform and by V-REP simulation software. The results show that the proposed Bayesian ProMPs-PI² learning method enables the robot to complete the generalization learning from demonstration tasks to unfamiliar tasks quickly and accurately, and acquire new skills successfully.