Pouring Sequence Prediction using Recurrent Neural Network

TL;DR

This paper presents an RNN-based approach to predict pouring sequences, including velocity and weight, to enable robots to imitate human pouring actions despite environmental variations.

Contribution

The study introduces a neural network model that learns and predicts complex pouring sequences, incorporating velocity and weight estimation, evaluated with DTW.

Findings

RNN effectively predicts unseen pouring sequences.

Model accurately estimates pouring velocity and container weight.

DTW demonstrates strong prediction performance.

Abstract

Human does their daily activity and cooking by teaching and imitating with the help of their vision and understanding of the difference between materials. Teaching a robot to do coking and daily work is difficult because of variation in environment, handling objects at different states etc. Pouring is a simple human daily life activity. In this paper, an approach to get pouring sequences were analyzed for determining the velocity of pouring and weight of the container. Then recurrent neural network (RNN) was used to build a neural network to learn that complex sequence and predict for unseen pouring sequences. Dynamic time warping (DTW) was used to evaluate the prediction performance of the trained model.