Continual learning of human-like arm postures

• Main Author: Gulletta, Gianpaolo
• Publication Date: 2021
• Other Authors: Erlhagen, Wolfram, Bicho, Estela
• Language: eng
• Source: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Download full: https://hdl.handle.net/1822/88042
• topic: Continual learning; Human-like robot motion; Human motor control; Developmental robotics; Warm-starting IPOPT; Human-like arm posture prediction; Cognitive robotics; Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática

Inspired from established human motor control theories, our HUMP algorithm plans upper-limb collisions-free movements for anthropomorphic systems, which show kinematic human-like features [1]. Related cognitive issues can be further resolved when robots act as they are familiar with their workspace and can take initiative faster than in the early onsets of a task. Here, a continual learning technique is proposed to improve the performance of the HUMP under uncertainties of the items in a given scenario. Given the locality of the optimization-based HUMP algorithm, a meaningful initial guess, predicted from similar past motion experiences, can significantly reduce the computational cost and put the robot into action arguably faster than in the first attempts of planning with inexperienced initial guesses. This prediction is proposed to be incrementally refined by an optimal locally weighted regression method that operates on datasets of situational features that are regularly updated as new movements are planned by the robot in similar scenarios. The proposed cyclic experiential learner is tested on the selection of optimal human-like target postures in a reaching task with a large obstacle obstructing the straight-line path towards a given target. Results demonstrate the capability of extracting meaningful situational features in few sessions of online learning with a very limited size of the datasets. Comparisons with simple Euclidean locally weighted regression and random initializations showed the capability of planning target configurations of better quality with less computational cost. The proposed approach also exhibits to be robust against the interferences of new incoming samples depicting slightly changed situations of the same task.