Computer Science and Engineering > Artificial intelligence and robotics


http://airlab.ws.dei.polimi.it/

Focus

The research group integrates the activities of a number of strictly interacting people, each one working on one or more research topics, roughly covering the general areas of Artificial Intelligence, Autonomous Agents, Computational Intelligence, Machine Learning, Autonomous Robotics, Computer Vision, and related philosophical aspects.

Artificial Intelligence
The research line is mainly working on topics related to autonomous agents and multi-agent systems. One of the issues of interest concerns the development of solution concepts and of cooperation mechanisms for the interaction of rational agents able to be deployed in different real-world situations.

Autonomous Robotics
To enable autonomous robots widespread application, one of the open problems concerns the ability to perform a task while satisfying market constraints such as reliability, safety, power self-sufficiency, and cost/effectiveness, among others. Robotics in the real world requires us to implement autonomy in terms of robustness, reliability, and uncertainty management in the whole path that goes from perception to action. Robust perception, rich world modelling, planning under uncertainty, interaction, and learning are the key aspects that we are facing. The research in robotics has also investigated methods and techniques for autonomous robot benchmarking and evaluation.

Computer Vision
Visual analysis of complex scenes, where the essence of the present objects cannot be captured by purely geometric models, is one of the most challenging problems in the present and future research. Given the relevant quantity of available mathematical tools, part of the fruitful effort can be aimed at identifying the subtle approximation/simplification that make a problem tractable and at generalizing to different setups. In the realm of complex scenes understanding, the next step will be to recognize what object the camera is pointed to, what is the purpose of the perceived movement, and what is going to happen then.

Machine Learning
Reinforcement learning was born as a bio-inspired approach to enable agents/robots to autonomously achieve complex behaviors in unknown environments. Currently, the field is reaching a more mature state and is providing effective solutions to many real-world applications that are too complex for other approaches.
In a more biologically-oriented perspective other kind of learning have gained importance in modeling neuronal architectures. In this case, learning often has a role in setting internal parameters of the robot, or sometimes in making associations in different spaces. In those cases the “one shot learning”, where an association is created without the need to explore the entire space, is the wanted target.

Philosophy of Artificial Intelligence and Robotics
In the last years attention has been focused on more specific issues emerging from the practice of AI and robotics (including ethical ones). Moreover, some discussions within AI and robotics have been generalized to the whole computer science and engineering field, giving birth to a new area of research called Philosophy of Computer Science.

Most relevant research achievements


Communication, norms, and strategy in multi-agent interaction
Strategic interactions in multi-agent systems involve aspects of strategy design, norm compliance, and communication. The strategy design when agents are rational is studied by resorting to tools from microeconomics and algorithmic theory. Agent communication is dealt within the context of a number of institutions, used to specify both the semantics of messages and the norms regulating the interaction.

Computational intelligence methods in the prediction of biological properties from chemical structure
We developed a family of hybrid system architectures, combining rules and soft computing, for toxicity prediction and drug discovery. We designed new data mining methods to take the structural information directly from SMILES for building statistical models. We set up the international web site http://vega-QSAR.eu, with developers and authorities in USA, UK, and Italy.

Computational intelligence and games
Our research focuses on the development of computational intelligence tools that support designers and developers to automatically generate high-quality content, based on the feedback provided by a panel of experts or users.

Genetic and evolutionary computation
We developed evolutionary methods to solve problems in a wide variety of application domains including machine learning, hardware-software co-design, and software test. We also investigated the mechanisms underlying artificial evolution and introduced methods to improve the current state of the art both in genetics-based machine learning and in numerical optimization.

Machine learning in the real world
The research group has a strong tradition in machine learning both from a theoretical and from a practical perspective which brought to the successful application of machine learning techniques to: natural resource management, landslide susceptibility analysis, brain-computer interfaces, intelligent transport systems, control of semi-active suspensions, sleep staging and analysis, affective interaction, user modelling, and robotics.

Localization, mapping, and exploration for mobile robots
A proper perception and modelling of the surrounding environment is a key component for autonomous robots. We have achieved significant advancements in autonomous map building and world modelling for single and multi-robot systems, leveraging on range sensing and vision. Moreover, a number of techniques for autonomous active exploration of environments have been proposed.

Autonomous robots
We have implemented several autonomous robots including a robotic wheelchair that can be driven with different interfaces (including miographic interfaces and brain-computer interface), autonomous robots to play games with people, robots for emotional interaction, all-terrain vehicles (in collaboration with MERLIN lab), and some autonomous flying drones.

Robots as (and for) biological systems
We developed humanoid robots incorporating a bio-inspired view of architecture and control. Some of our systems use a pure neural approach to cognitive architectures. We also studied the human neural and neuromuscular organization to design devices able to mimic human cognition and control. The applications are mainly in rehabilitation robotics and prostheses, where we developed an EMG-based control.

Calibration, localization, and odometry using catadioptric cameras
The study of new geometric properties of non-central catadioptric cameras led to techniques to localize straight lines and spheres from single calibrated catadioptric images, as well as a technique for visual odometry and ground plane reconstruction from uncalibrated images.

Foundational issues in computer science and engineering
The research group has investigated some foundational issues in computer science and engineering with particular attention to the status and role of experiments and computer simulations used as experiments and to the relationships between computer engineering and philosophy.

Autonomous Robots Benchmarking and Standardization
The research group has participated from its foundation to the Euron Special Interest Group on Good Experimental Methodologies and Benchmarking; we participated to the writing of the guidelines for scientific paper reviews. We have coordinated (FP6 – RAWSEEDS http://www.rawseeds.org), and we participate (FP7 – RoCKIn http://rockinrobotchallenge.eu/) to EU projects on the development of methods and techniques for autonomous robot benchmarking. We also participate to the IEEE standardization effort on ontologies and map representation for autonomous robots.