|CASALINO ANDREA||Cycle: XXXII |
Section: Systems and Control
Tutor: BASCETTA LUCA Major Research topic
:Toward a real human-robot collaboration in industrial cell, using predictive algorithms and optimal scheduling strategies
Advisor: ROCCO PAOLOAbstract:
One of the most active contemporary research topic in robotics, is the study of methodologies and techniques designed to let industrial manipulators work side by side to human, in plants where humans and robots collaborate to complete common tasks. In these kind of plants, operators are required to execute high cognitive tasks, like e.g. assembly operations that could be too much difficult to fully automatize, while industrial robot assist and interact with human in many ways or keep on doing some other assigned activities. The combination of human flexibility and machine efficiency can essentially reduce the amount of fixed production costs and therefore the Small and Medium sized Enterprises (SMEs) are the ones that could benefit the most by the introduction of the so called “CoRobots”. In a not distant future industrial scenario, robots will share a common workspace with humans, with the aim of having a real strong collaborative interaction. The field of Human Robot Interaction (HRI) has interested lots of researchers in the recent years. Many efforts were devoted to develop algorithms and methodologies to let human and robot coexist in a common environment. In a typical HRI control architecture it is essential to insert a block dedicated to manage data provided by multiple sensors that acquire information concerning the environment. Detecting the presence of humans in the scene it’s clearly a key aspect, but also predicting their intention of motion is important as well. In the literature it is possible to find a lot of works investigating this aspect, even though it is difficult to find a general methodology devoted to understand human intention of motion, in a completely unstructured environment and without knowing a finite set of the possible goals that a human could reach. So this problem could be still considered as an open one. Another feature of a classical HRI controller is to guarantee a certain safety level. This goal is usually obtained designing reactive behaviours for the robot, that are activated in unexpected situations that could arise to avoid collisions between human and robot. Most of the algorithms developed in the field of HRI solve efficiently the problem to enable a robot and a human to carry on disjoint activities in a common scene, without incur in unsafe situations. By the way, the intent of this thesis is to investigate more sophisticated techniques to shift the interaction of human and robot from a general coexistence to a real collaboration (HRC). The aim is in particular to develop algorithms and methodologies to let an industrial robot reach more elaborated cognitive abilities, in order to be able, for example, to dynamically plan joint tasks with a human agent. Aspects related to safety between human and robot will be considered too. These thesis will focus on two main goals:
- Address the problem of recognize the current action a human is executing at a first stage, while at a second forecast the possible future sequence of actions. It’s clear that for the first case it is necessary to find a way to manage the stochasticity nature of human motion, since the same action (for example take an object from a box) can be reproduced with slightly different trajectories. Therefore to infer the goal a human is intended to reach, when can retrive a set of observations from sensors perceiving the scene, able to give during time the position of some points of the human silhoutte, the gaze direction and other. These collection of indirect observations are used to feed powerfull statistical tools like HMM (Hidden Markov Model) or more more general Bayesian Network, to make inference about the current human action. The results of the first stage prediction are used even for a second stage prediction, in which the objective is to forecast more in general the future sequence of human actions. Again it is possible to tackle this problem with a statistical perspective, exploiting the methods used in the field of pattern recognition.
- Develop algorithms to let an industrial robot autonomously program its present and future activities considering the aforementioned human’s actions recognition and prediction stage. In this case the problem is essentially to schedule the activities that a robot will undertake, to be as much as possible compliant with those forecasted for the human. This will make a robot able to predict when to assist a human agent with the most appropriate action. On the other hand, for those temporal windows unlikely to require human robot collaboration, the robot must select the proper autonomus sequence of actions, to realize an overall feasable schedule (that is, not arrive late to the collaborative tasks). Since in principle many autonomus actions will be available, the problem became to find the optimal schedule. To solve this control problem, the powerfull framework of Petri Net will be exploited. One last aspect is to execute the scheduled plan of robotic actions with the aim of promote a safe collaboration, computing the appropriate trajectories for the manipulator. This became a non classical path planning problem, since the human in the scene is not a fixed obstacle, but his volume of occupancy during time can be estimated with a certain level of uncertainty according to the sequence of actions forecasted.