Learning from Demonstration with Gaussian Process Approach for an Omni-directional Mobile Robot

García, Daniel; Gorrostieta, Efren; Soto, Emilio Vargas; Rivero, Cristian Rodríguez; Delgado, Guillermo Diaz

doi:10.1109/tla.2018.8362164

Cited by 1 publication

(1 citation statement)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Both approaches require an effective mapping between sensors and actuators, and this can be done by statistical analysis of the data collected during demonstrations and by adopting non-linear regression techniques. Common methods include Gaussian Processes Garcia et al (2018), Artificial Neural Networks (ANN) Pomerleau (1989), Radial Basys Function ANN Gribovskaya et al (2011) Kulic and Vukic (2005), convolutional ANNs Zhang et al (2018), Support Vector Machines Khansari-Zadeh and Billard (2010), Fuzzy systems Teng et al (2023), Decision trees Sheh et al (2011) Bain and Sammut (1999b), among others.…”

Section: Introductionmentioning

confidence: 99%

Learning from Demonstration in Embedded Hardware for the Composition of Micro-Skills in Mobile Robots

Triana,

Arboleda,

de Oliveira

2024

Preprint

View full text Add to dashboard Cite

Designing mobile robot controllers for non-expert users is a complex task that involves programming abilities and developing accurate mathematical models for representing the robot's kinematics, sensor observations, and parameter estimations. Learning from demonstration (LfD) is a technique that enables robots to autonomously learn and perform new tasks from the observations of human demonstrations. Obtaining models with low computational complexity is of utmost importance for embedded robotics systems. This paper proposes the usage of LfD for learning three simple micro-skills (move forward, turn clockwise, and turn counterclockwise) that can be combined to efficiently perform more complex skills on a mobile robot. An adaptive single-layer perceptron neural network (SLP) and a Particle Swarm Optimization (PSO) training algorithm were implemented using a low-cost System-on-chip (SoC) device, allowing the robot to learn the wheel's speed profile for each demonstrated micro-skill and compose them to navigate around obstacles in complex scenarios. Real-world scenarios were used to statistically analyze the robustness of the proposed methodology, achieving a success rate of 100% for the known scenarios and a success rate of 61.81% for the unknown scenarios. Experimental results demonstrated that the robot correctly moves on several unknown scenarios that compose the three taught micro-skills, avoiding obstacles and correcting deviations from the initial position.

show abstract