#	Pagina
attuale pagina	/open-h2020/projects/195745/results.html

Report

Teaser, summary, work performed and final results

Periodic Reporting for period 1 - LEACON (LEArning-CONtrol tight interaction: a novel approach to robust execution of mobile manipulation tasks)

Teaser

Summary

One of the main challenges of roboticists in both academia and industry is to take robots out of the factories and let them enter into unstructured environments, such as houses, hospitals, small manufacturers and dangerous area.
The scientific objective of the project is to take a step towards the presence of robots in such environments.
Currently, there are still important obstacles to the massive diffusion of robotic systems in the fields described above. First of all, programming robots with the classical methods is still too expensive and time-consuming due to intrinsic complexity of manipulation tasks.
A second limitation is that planning the robot motion completely off-line may likely bring to a failure of the assigned task, since a high degree of uncertainty is present. Moreover, robots that work in anthropic environments should be able to understand and classify human actions and behaviour for an effective human-robot interaction.
In order to tackle these limitations, the LEACON project has the objective to develop a framework that:
-allows combining learning and perception-based control to achieve higher robustness
-exploits multimodal and crossmodal perception (tactile, proximity, visual, force sensors) to increase the robustness to unforeseen events
-allows the robot to recognize human actions
A first, shorter-term benefit of having robots with learning capabilities is in agile manufacturing. In such applications, we need flexible production processes and low programming costs. With those technologies, process experts with very limited robot programming skills could re-adapt the assembly line in an faster and much cheaper way. Also, machine laerning for robotics is an enabling technology for applications of the future, such as hospital automation, home automation, automous farming, operations dangerous for humans and human-robot collaboration.

Work performed

\"LEACON contributes to the advancement of state of the art in robotics within four main aspect:
1. A strategy to combine learning with reactive control, in order for robots to learn in a more robust way while avoiding irreversible events.
This research has been accepted both to ICRA (IEEE International Conference on Robotics and Automation) 2018 and Robotics and Automation Letters (RA-L).

2. The concept of cross-modal visuo-tactile perception has been introduced. Cross-modal perception allows a robot to gather knowledge with a sensing modality like vision, and to exploit such a knowledge using a second modality like touch. This way, the robot is able. for example, to gather knowledge of objects using vision, and, later, to recognize such objects by using only touch, without touching any objects before. This approach has to potential to enhance the flexibility of robot systems when working in unstructured environments, where the ability to change sensing modality is essential. The work has been accepted at ICRA 2017 and a journal extension with the latest is now preparation for IEEE Transactions on Robotics.

3. Two time and data-efficient approaches for recognition of human actions have been developed, called CODE and FADE. Using those approaches robots are able to recognize human actions.
CODE (COordination-based action DEscriptor) is based on a novel way to consider a similarity between two human whole-body actions based on how humans coordinate their body parts, while performing an action.
The CODE approach has been published in IEEE Robotics and Automation Letters (RA-L) and invited for presentation at ICRA 2017.
FADE, on the other hand, is a frequency-based approach. It leverages the properties that human motion cannot have significant frequency components higher than 15 Hz.
FADE has been published in Autonomous Robots and at the International Conference on Intelligent Robots and Systems (IROS 2016).

4. An approach for learning based on trial and error, PI-REM, has been developed. It aims at increasing the data efficiency leveraging approximated prior knowledge.
The work has been published at IROS 2017 and a journal paper is now in preparation.

The work carried out within LEACON has been published in important journals (RA-L, Autonomous Robots) and conferences (ICRA and IROS).

A workshop and a special issue were organized to further discuss the topic within the scientific community. During the \"\"tag der fakultÃ¤t\"\" at the Technical University of Munich, the topics of LEACON were disseminated outside the scientific community.

The resaercher designed and was lecturer of the class \"\"Reinforcement Learning for Robotics\"\", which is based on the research activities of LEACON.
It is one of first classes worldwide that covers theory of trajectory-oriented reinforcement laerning applied to robotics, emphasizing the connection between learning and optimal control.
Due to its success, the class will be given also the next years within Prof. LeeÂ´s chair.

As further dissemination activities, the researcher organized with Dr. Fanny Ficuciello and Dr. Sylvain Calinon the workshop: \"\"Learning and Control for Autonomous Manipulation Systems: the Role of Dimensionality Reduction\"\" held at ICRA 2017 and he was guest editor of the RA-L special issue associated to the workshop.

Thanks to the research results achieved within LEACON, Dr. Falco was offered and accepted a position at ABB, Corporate Research in VÃ¤sterÃ¥s, Sweden as a tenured senior scientist and project manager.
Therefore, the methodologies and skill acquired in LEACON have a direct impact both on researcherÂ´s career and on the robots of the future both in flexible manufacturing and service robotics.
\"

Final results

To summazie, the progress with respect to the state of the are the following:
- a strategy for synergy between a learning and reactive control layer was proposed. This way, robot will be able to learn in a rubust way to irreversible events.
- the concept of cross-modal visuo-tactile perception was introduced for the first time in the robotics community and a pilot strategy to tackle this problem was proposed.
- two strategy for human action recognition (CODE and FADE) are introduced, which allows a robot to recognize human actions in a time and memory efficient fashion. Those approaches do not need massive datasets, powerful GPUs and complex classification strategy. Therefore they are suitable for new generation service robotics systems that have to be efficient in terms of power consumption and cumputational cost.

The methods developed in LEACON are a step towards robotics in both flexible manufacturing and service robotics in unstructured environments.
In agile manufacturing, the robot with learning capabilities will be programmed by users who are not expert in programming and will allow to change to production line very quickly. Also, the methods developed in LEACON will contribute in the use of cobots able to recognize human actions and to refine their skills by trial and error.
Learning capabilities will enable robots to be applied in new areas such as home automation, automatic farming, hospital automation, missions in dangerous areas.