In this four year project, led by Birmingham, and involving six universities and 35 researchers, we developed algorithms and software to enable robots to monitor their own knowledge and plan with this. The result was a series of robot systems. Dora mapped her world, planning to acquire new knowledge as necessary to perform the tasks she had. George conversed with a human to find out more about the world. Dexter grasped and manipulated objects. Dora and George both plan their activity in order to acquire new information, representing what they know, what they don’t know, and how what they know will change as they act.

Participants: Jeremy Wyatt, Nick Hawes, Aaron Sloman, Richard Dearden

More information can be found on the CogX project website.

In this three year project we worked on object manipulation tasks. In order to work naturally in human environments such as offices and homes, robots of the future will need to be much more flexible and robust in the face of novelty than those of today. In GeRT we will develop new methods to cope with novelty in manipulation tasks.

Humans cope so seamlessly with novel objects that we do not think of grasping a new cup, or screwing the lid off a jar we haven't seen before as challenging. But this kind of everyday novelty in manipulation tasks is hard for a robot. Currently the most advanced robots can perform a task such as making a drink, which involves grasping, pouring, and twisting off a cap from a jar. But the rules for how to pick up every single object must be programmed. So the ability to manipulate fifty different objects means writing fifty different programs. Also when the robot encounters an object it hasn't seen before, it can't grasp the object. Even worse than this is the fact that if one object in a task changes then the program for the whole task may need to be rewritten. This is because how we manipulate an object depends on the task, and the other objects involved. Suppose, for example, that we substituted a mug for a glass in a task that involved pouring liquid from the mug or glass into another object. The pouring position would change, as would the grasping position. Perhaps we would grasp the mug by the handle, and then tip it sideways to pour.

All of this means that if robots are ever going to be useful in natural settings where manipulation is involved that they need ways of generalising on the fly to cope with novel objects, and perhaps novel tasks. Our approach is to take a small set of existing robot programs, for a certain robot manipulation task, such as serving a drink and to give the robot the ability to adapt them to a novel version of the task. These programs constitute a database of prototypes representing that class of task. When confronted with a novel instance of the same task the robot needs to establish appropriate correspondences between objects and actions in the prototypes and their counterparts in the novel scenario. In this way the robot can solve a task that is physically substantially different but similar at an abstract level. To achieve this we will use a variety of techniques from machine perception, machine learning, and artificial intelligence techniques such as automated planning.

More information is available at the GeRT project website.

Participants:Jeremy Wyatt , Rustam Stolkin, Richard Dearden

In this three year project, led by Birmingham, we are developing new methods for object perception, representation and manipulation so that a robot is able to robustly grasp objects even when those objects are unfamiliar, and even though the robot has unreliable perception and action.

As an example imagine a robot that must load a dishwasher with cookware and crockery. Some of the object instances in the task are familiar to the robot, some are novel in the sense that they are unfamiliar instances drawn from familiar object categories. For example, perhaps the robot has previously manipulated mugs, but not a water jug. This new object has a handle, and so the robot can try to adapt a grasp it already knows for the handle of the mug to that of the jug.

Our approach assumes that objects are made of parts, and those parts of smaller parts. We have already shown that this approach enables much more robust computer vision systems that can reliably recognise objects at the category level (dogs, cars, faces, bicycles etc). We are now applying it to object manipulation. We are also studying how the robot’s fingers and cameras actively gather information to help manipulation.

More information is available at the PacMan project website. 

Participants: Jeremy Wyatt, Ales Leonardis

STRANDS aims to enable a robot to achieve robust and intelligent behaviour in human environments through adaptation to, and the exploitation of, long-term experience. Our approach is based on understanding 3D space and how it changes over time, from milliseconds to months. We are developing novel approaches to extract spatio-temporal structure from sensor data gathered during months of autonomous operation. Extracted structure will include reoccurring 3D shapes, objects, people, and models of activity. We will also develop control mechanisms which exploit these structures to yield adaptive behaviour in highly demanding, real-world security and care scenarios.

By autonomously modelling spatio-temporal dynamics, our robots will be able run for significantly longer than current systems (at least 120 days by the end of the project). Long runtimes provide previously unattainable opportunities for a robot to learn about its world. We are integrating our advances into complete robot systems to be deployed and evaluated at two end-user sites: a care home for the elderly in Austria, and an office environment patrolled by a security firm in the UK. The tasks these systems will perform are impossible without long-term adaptation to spatio-temporal dynamics.

STRANDS will produce a wide variety for results, from software components to an evaluation of robot assistants for care staff. These results will benefit society in a range of ways: researchers will be able to access our results as open-access papers, software and data; our methodology for creating long-running robots will encourage roboticists to tackle this unsolved problem in our field; industrialists will see how cognitive robots can play a key role in their businesses, and access prototypes for their own use; and society will benefit as robots become more capable of assisting humans, a necessary advance due to, for example, the demographic shifts in the health industry.

Participants: Nick Hawes, Jeremy Wyatt

More information is available at the Strands project website.

The aim of CoDyCo is to advance the current control and cognitive understanding about robust, goal- directed whole-body motion interaction with multiple contacts. The CoDyCo project is a four-years long project and starts in March 2013. At the end of each year a scenario will be used to validate our theoretical advances on the iCub Robot.

Methodologically, CoDyCo activities are divided into four categories, deeply intertwined: control theory, machine learning, human behavioural experiments and software development. Each of them serve CoDyCo’s objectives to: 1) develop a general software toolkit for whole-body dynamics computation with multiple external contacts, 2) conduct human behavioural studies for understanding human use of external contact, including interpersonal cooperative contacts in natural whole body tasks, 3) develop a control architecture for whole body coordination and regulation of whole body compliance, and 4) leverage machine learning methods for acquiring models of compliant contact with the environment and physical interactions with humans.

At Birmingham we are studying human behaviour and developing new methods for compliant control.

Participants: Michael Mistry, Chris Miall, Alan Wing

More information is available at the CoDyCo project website.

Sensory information is often temporally structured and our brain exploits the regularity for perception. We study how expectations can influence our perception of rhythm, music, and timing of stimuli.

The state of activity in the brain fluctuates regularly. The characteristic of the oscillations and their synchronization is related to external stimulation and cognitive processing. Using physiological recordings we investigate the parameters and predictability of neurocognitive models.

The properties of the environment as well as our internal states can influence how we move. We record repetitive actions as in dance and music production to investigate what are the stimulus properties, cognitive processing, and muscular actuation that allow us to synchronize to external rhythms.

Principal Investigators

Dr Max Di Luca
Dr Simon Hanslmayr
Professor Alan Wing 
Professor Kimron Shapiro
Professor Jane Raymond
Dr Maria Wimber

Funding

Marie Curie ‘TICS’

Focus

Temporal and Rhythmic Perception. Neural Oscillations. Sensorimotor Synchronization.

Techniques

Psychophysics to measure perception
Electroencephalography (EEG) to look at internal oscillations Transcranial stimulation (tDCS, tACS, TMS) to perturb brain rhythm Motion Capture to record body movements Computational Modeling to integrate the three streams of research

This work uses qualitative and probabilistic models to learn then exploit the structure and predictable dynamics of the appearance, configuration and human activities present in every day environments. Examples include learning the typical spatial arrangements of objects on tabletops over time to support robot object search, or the fastest routes for a mobile robot to take through an environment at a particular time of day.

Principal Investigators:

Dr Nick Hawes
Dr Jeremy Wyatt

Funding:

EU, EPSRC

Focus:

Learning semantic structure from human environments.

Techniques:

Mobile robots; probabilistic and/or qualitative modelling of space, time, activity or function; machine learning; AI planning; computational vision.

Summary:

A long term project attempting to understand the mechanisms and intermediate stages in evolution of biological information processing, including precursors of human intelligence.
Further information:

www.cs.bham.ac.uk/research/projects/cogaff/misc/meta-morphogenesis.html 

Investigators:

Professor Aaron Sloman

Funding:

Unfunded. There are no plans to apply for funding.

Focus:

Attempt to understand biological intelligent systems and possible future intelligent robots with similar capabilities, by investigating the major transitions in biological information processing from pre-biotic materials onwards. The transitions include changes in opportunities and problems posed by the environment, changes in sensory-motor morphology, changes in information content aquired, stored, derived, and used, changes in forms of representation, changes in mechanisms for manipulating information, changes in forms of information and mechanisms used in reproduction, changes in forms of learning, both across generations and within individuals, and changes in social/cultural forms of information processing and information transmission, including teaching and education.

Techniques:

Wild speculation tamed by philosophical analysis, study of literature, critical analysis and, when possible, experiments on humans and other animals. Eventually construction of working intelligent systems will be required, when ideas become sufficiently precise. There is already a vast amount of relevant literature, from many disciplines, including ethology, developmental psychology,  linguistics, AI, Robotics and philosophy.

Summary:

This research project is part of a project aiming to develop a novel robot-mediated training method to improve reaching movements of stroke patients. In the training task, the patient will hold the robotic arm that will allow both recording of movement parameters and providing assistive and/or resistive forces to guide the impaired movement. The novel adaptive algorithm applies a selects next training movements based on recent movement performance parameters and a general learning principle. Additional objectives of this research include thorough characterization of upper-limb reaching performance and learning in healthy and stroke populations and evaluation of recovery-related neural reorganization.

Robotic devices map upper limb performance and adaptively schedule rehabilitation

Investigators:

Professor Chris Miall
Professor Alan Wing
Dr Orna Rosenthal
Dr Jeremy Wyatt

Funding:

MRC

Focus:

Optimizing motor learning and recovery (of the upper-limb) based on general learning principles

Robot-mediated motor training

Kinetic (spatiotemporal) and kinematic (muscle-joint) aspects of reaching movements across the work space

Mechanisms underlying motor recovery

Techniques:

Psychophysics, kinetic analysis, robotics, EMG (muscle activity), motion analysis, MRI

Summary:

Using data from experiments in perception, decision making and learning we are developing advanced statistical models to compare human behaviour to optimal Bayesian observer models. We compare these models to data from brain imaging to learn how the brain performs these computations.

Principal Investigators:

Dr Ulrik Beierholm
Dr Massimiliano Di Luca
Professor Andrew Howes
Professor Uta Noppeney

Funding:

EU, Marie Curie

Focus:

Models are being developed to improve our understanding of how the brain performs multi-sensory integration, perception, learning, and economic decision making.

Techniques:

Statistical model, Reinforcement learning models, psychophysics, EEG, fMRI

Summary:

This research project aims at improving object detection by optimally combining the complementary strengths of human and machine perception. The basic idea is to provide machines not only with environmental inputs for object detection, but also with neural activity features (i.e. EEG signals) recorded in humans exposed to exactly those sensory signals during object detection. In the second step we will train humans to enhance these neural activity features and gate them into awareness to improve human object detection performance.

Principal Investigators: 

Professor Ales Leonardis
Professor Uta Noppeney

Funding:

EU

Focus:

Understanding and enhancing cognition and performance;

Exploring the role of cognitive neuroscience in understanding, managing and optimizing human performance. 

Techniques:

Computational Vision, Machine Learning, EEG

Summary:

Psychophysics, neuroimaging and computational modelling are combined to investigate how the human brain integrates signals from multiple senses into a coherent percept of our environment. Computational operations are described using models of Bayesian inference and learning. Neural mechanisms are characterized by combining multimodal imaging (fMRI, EEG, TMS) with advanced neuroimaging analyses including multivariate decoding (e.g. support vector machines) and effective connectivity analyses (e.g. Dynamic Causal Modelling).

Investigators:

Dr Ulrik Beierholm
Dr Massimiliano Di Luca
Dr Robert Gray
Professor Uta Noppeney
Professor Alan Wing

Funding:

ERC, EU, Marie Curie, BBSRC

Focus:

Understanding the computational operations and neural mechanisms of information integration

Techniques:

Psychophysics, multimodal imaging (fMRI, EEG, TMS), computational modelling (e.g. Bayesian inference)

Summary:

Haptics (=active touch) is the combination of tactile and proprioceptive information during contact and manipulation. The haptic modality is involved in both perceiving the environment and manipulating it physically. The influence of sensory information on action and the reverse influence of active movement on perception are far from being understood. Our research in the field of haptics thus aims at building computational models of human behaviour and perception. Such models can be utilized in technical applications (i.e., robotic systems, force-feedback interfaces, telepresence, reabilitation devices).

Investigators:

Dr Massimiliano Di Luca
Dr Robert Gray
Dr Hoi Fei Kwok
Dr Roberta Roberts
Professor Chris Miall
Professor Alan Wing
Professor Jeremy Wyatt

Funding:

The European Commission, EPSRC, NERC, The Royal Society, The Leverhulme Trust, The British Council, The Department of Trade and Industry, AWM, DAAD

Focus:

Role of active exploration in perception
Trade-off between exploration and exploitation
Rehabilitation for patients with motor deficits
Learning from human grasping to improve robotic grasping

Techniques:

Psychophysics, motion analysis, robotics, computational modelling

Summary:

Developing clinically applicable techniques for cognitive screening after brain injury, examining cognitive deficits in relation to brain lesion, using fMRI to develop neural-level models of neurorehabilitation.

Principal Investigators:

Dr P Rotshtein

Funding:

MRC and Stroke Association

Focus:

Predicting neural recovery from multi-modal brain imaging Techniques: Neuropsychology, MRI (structural and functional), computational modelling

Summary:

Using advanced pattern classifier systems to distinguish the neural response to different depth cues to assess whether depth information is coded in a common manner. Also uses psychophysical work to characterise how depth cues may be analysed in computer vision and used to create graphic surfaces.

Principal Investigators:

Dr P Tino

Funding:

BBSRC, EPSRC, EU

Focus:

Understanding the neural basis of depth perception

Techniques:

Psychophysics, fMRI, computational modelling

Summary:

Using fMRI with neuropsychological patients and combined with TMS to show the necessary role of different regions with cortical networks, and using computational models to define the functional roles of different regions.

Principal Investigators:

Dr C Mevorach

Professor U Noppeney

Funding:

BBSRC, MRC

Focus:

Using neuropsychological-fMRI and fMRI-TMS to decompose interacting neural networks, computational modelling.

Summary:

The role of the cerebellum in forward planning of action studied through robotic systems where force feedback is varied.

Principal Investigators:

Professor C Miall, Professor A Wing, Dr J Wyatt

Funding:

BBSRC, Wellcome, HFSPO, EU, DTi

Focus:

Perception and action

Techniques:

Human psychophysics, robotics, fMRI

Summary:

Using and developing multi-modal imaging techniques and multi-voxel classifier systems to understand individual differences in perception and learning, and using this to guide procedures for maintaining plasticity across the age ranges. We are also modelling human action.

Principal Investigators:

Professor Uta Noppeney, Dr A Bagshaw, Professor A Wing

Funding:

BBSRC, EPSRC, ESRC

Focus:

Understanding and exploiting neural plasticity

Techniques:

Multi-modal imaging, human psychophysics, computational modelling

Summary:

We are developing procedures for multi-modal imaging, combining the time course of EEG with spatial resolution of MRI, and combined MRI and Trans-cranial Magnetic Stimulation to provide improved analyses of neurological disorders and ageing.

Principal Investigators:

Dr T Arvanitis

Dr A Bagshaw

Professor Uta Noppeney

Funding:

BBSRC, EPSRC, EU

Focus:

Developing techniques for multi-model imaging, combining fMRI, EEG, diffusion tensor imaging and magnetic resonance spectroscopy (MRS), applied to understanding cognition after brain injury, epilepsy, in adults and children. 

How are tools (by which I mean physical objects) used in creative practices (like painting, drawing, sculpting)? I am interested in what creative practitioners do with their tools and how this practice relates to cognition.  A recent paper raised some of the issues of interest (https://link.springer.com/article/10.1007/s13347-017-0292-0). There are several types of project that could be considered in relation to this:

• Data from sensors (on tools or on the person) allow us to characterise discrete actions, but how can we better understand sequences of movements? In other words, how can time-series analysis help understand complex movement?

• Is there a relationship between cognitive aspects of creativity and physical actions (as implied in the paper)? In other words, what does an embodied theory of creativity look like?

• Do different creative practices have similar underlying phenomena or are there distinct phenomena surrounding each practice?

Techniques:

The projects could involve some of the following (but we can work around most of these if you don’t have the prior knowledge): Signal processing (with R or MatLab), Basic electronics (making things with sensors and Arduinos), Data collection in the field (at the workplaces of practising artists), Phenomenological interviews (or other form of in-depth interviewing to gain insight into experiential aspects of creative practice).

Investigator:

Chris Baber

The ability to draw inferences from a collection of disparate information sources is key to many intelligence analysis professions.  I’ve been looking into how ‘uniformed’ professions might go about making sense of complex data.  Here’s a fairly recent piece of work on this: https://dl.acm.org/citation.cfm?id=2870212. Projects that relate to this area could involve the following:

• How do people create a ‘gist’ (or working model) of inferences when dealing with lots of different sources of data?  Does this involve the creation of narratives (as work on jury decision making suggests) or does it involve the creation of diagrams (or sketches)?
• How do people challenge their own (and other people’s) inferences during sensemaking?
• How can technology help or hinder people’s abilities to make sense of multiple information sources?

Techniques:

It would be quite nice to make a software prototype for this (using one of the many Javascript toolkits, like d3), but it is equally possible to make a paper-based set of material for experiments (this is what I’ve done in the North x Southwest Exercise that is reported in the paper linked above).  It would be good to run an experiment, with individuals or small groups working together.  I’d quite like to run a study using Agent-Based Models to see how information sharing has an impact on sensemaking.

Investigator:

Chris Baber

We have shown in several recent publications that participants’ ability to learn a new motor action through reinforcement-based (reward) feedback is highly variable. Specifically, while 66% of people are able to learn, 33% show a complete inability to learn. Our goal moving forward is to examine the difference between learners vs. non-learners in more detail and across a range of tasks; simple reaching, force-field adaptation, effort production and chord (similar to learning the piano) learning. With help from people in the lab, this will involve programming a novel task using one of several possible pieces of equipment (KINARM robotic manipulandum, motion capturing equipment, force capturing equipment) with either Matlab or Simulink (graphical Matlab). It will be expected that participants will be recruited and tested, then analysis will be performed on this data from basic kinematic analysis to developing computational models.   

Relevant publications:

• The contribution of explicit processes to reinforcement-based motor learning.
  Holland PJ, Codol O, Galea JM.
  J Neurophysiol. 2018 Mar 14. doi: 10.1152/jn.00901.2017. [Epub ahead of print]
• Predicting explorative motor learning using decision-making and motor noise.
  Chen X, Mohr K, Galea JM.
  PLoS Comput Biol. 2017 Apr 24;13(4):e1005503. doi: 10.1371/journal.pcbi.1005503. eCollection 2017 Apr.

Techniques

Matlab, Simulink

• Agent-based modelling of human social behaviour
• Cognitive Neuroscience with LEGO Mindstorms NXT
• Modelling of EEG-data
• What is wrong with Deep Neural Networks?
• Computational modelling of tool innovation

Agent-based modelling of human social behaviour

Summary:

An agent-based model (ABM) consists of independent agents interacting with each other and the environment. The behaviour of each agent is defined by a set of rules, only prominent to each agent. This type of model allows the researcher to observe how individual behaviours of agents (microscopic level) result in patterns on the macroscopic level (emergent behaviour). These properties make ABMs an ideal approach to modelling human social behaviour. For instance, the seminal Schelling model showed that social segregation can emerge from a minor preference to live in a neighbourhood populated by a similar social group. In collaborations with my colleagues from the social psychology group I aim to apply ABMs to a broad range of social phenomena, e.g. help seeking of ethnic minorities, social destigmatisation, social housing market, etc.

Principal Investigators:

Dr Dietmar Heinke

Funding:

EPSRC, ESRC, BBSRC

Techniques:

MatLab-programming

Cognitive Neuroscience with LEGO Mindstorms NXT

Summary:

In a recent publication we showed that LEGO Mindstorms NXT can contribute to our understanding of the brain, especially the interactions between visual cognition and motor behaviour. Based on this work numerous projects are possible: 

• There is a wealth of data on human reaching and grasping. However, these data have never been modelled in such a framework.
• In recent years, more and more evidence supports the idea that humans utilize feed-forward models (well-known from technical control theory) to guide reaching and grasping. Our Lego robot constitutes an ideal environment to test this theory in a natural set-up.
• Recent progress in developmental psychology shows that toddlers' interactions with the environment foster their learning progression. With the LEGO robot, we can mimic these interactions (developmental robotics) and for instance, explore how the recognition of objects can be learnt in such a scenario.

Investigator:

Dietmar Heinke

Techniques:

JAVA, Lego Mindstroms NXT, Image processing.

Modelling of EEG-data

Contemporary EEG techniques focus on a range of statistical methods, such as fitting general linear models, correlations, time-frequency analysis, dynamic causal models, etc. However, there has been little work on connecting EEG signals directly to computational models of cognitive abilities. Such models are similar to methods in artificial intelligence (AI) which aim at mimicking human cognitive abilities (e.g. playing and winning the ancient board game, GO). The aim of this project is to develop a novel method that tests such computational models by benchmarking them against EEG signals. In other words, this novel method will allow us to establish more directly links between human cognition and EEG signals than contemporary EEG methods. Consequently, we will be able to advance our understanding of neural mechanisms underlying cognitive abilities.

This project will focus on evidence from visual search tasks as test case for this novel approach. Visual search tasks aim to understand how humans analyse complex visual scenes and find behaviourally relevant object in them.

Techniques

Behavioural experiments, EEG experiments, MatLab, Computational modelling

Investigator:

Dietmar Heinke, Dominic Standage

References

Narbutas, V., Lin, Y.-S., Kristan, M., & Heinke, D. (2017) Serial versus parallel search: A model comparison approach based on reaction time distributions. Visual Cognition, 1-3, 306-325. https://doi.org/10.1080/13506285.2017.1352055

Lin, Y., Heinke, D. & Humphreys, G. W. (2015) Modeling visual search using three-parameter probability functions in a hierarchical Bayesian framework. Attention, Perception, & Psychophysics, 77, 3, 985-1010.

Heinke, D. & Backhaus, A. (2011) Modeling visual search with the Selective Attention for Identification model (VS-SAIM): A novel explanation for visual search asymmetries. Cognitive Computation, 3(1), 185-205.

What is wrong with Deep Neural Networks?

Deep neural networks (DNNs) have been behind recent headline-grabbing successes for artificial intelligence, such as AlphaGo beating the world's best player in the board game Go. Nevertheless, in many areas DNNs have not lead to technical solutions that match human abilities, especially with regards to object recognition. Perhaps the most obvious reason for such failures stems from the fact that DNNs employ very different processing mechanisms to those found in humans (e.g. Lake et al. 2016).

This project aims to compare DNN’s abilities with human abilities and ascertain how the two differ. Based on this understand we aim to develop novel approaches to automated object recognition.

Techniques

Tensorflow, Python

Investigator:

Dietmar Heinke

References

Lake, B., Ullman, T., Tenenbaum, J., & Gershman, S. (2016). Behavioral and Brain Sciences, 1-101.

Computational modelling of tool innovation

Humans’ ability to use tools has drastically transformed our planet, and is a skill we use in our daily lives. Building autonomous agents that possess this ability is crucial in allowing them to function in the world we live in. Recent developments in machine learning, such as deep reinforcement learning (DRL), has resulted in numerous successes in creating software and robot agents that are able to effectively use man-made tools. However, transferring the ability to innovate, such as inventing new tools or new uses for existing tools, has received relatively little attention so far in artificial intelligence (AI) research.

This project aims to develop an algorithm for inventing tools in the very simple scenario of the LEGO task. The LEGO task uses pairs of novel objects built with LEGO blocks. Participants are asked to determine how one object can transport (either lift or shift) the other object. They had three minutes to complete the task (see https://tinyurl.com/yck2rgfj and https://tinyurl.com/y9ug8snt for two examples of videos. The aim of the project is mimic human behaviour in these tasks.

Techniques

MatLab

Investigator:

Dietmar Heinke

References

Osiurak, F. (2014). What neuropsychology tells us about human tool use? The four constraints theory (4CT): mechanics, space, time, and effort. Neuropsychol. Rev., 24 (2), 88-115.

Osiurak, F., & Heinke, D. (2018) Looking for Intoolligence: A unified framework for the cognitive study of human tool use and technology. American Psychologist, 73(2), 169-185. http://dx.doi.org/10.1037/amp0000162.

• Tell me why: Explainable Representation and Reasoning in Robotics
• Theory of Intentions and Affordances for Assistive Robots

Tell me why: Explainable Representation and Reasoning in Robotics

Mobile robots assisting humans in homes or warehouses receive different descriptions of incomplete domain knowledge and uncertainty. Information about the domain often includes commonsense knowledge (e.g., "textbooks are usually in the library"), and quantitative descriptions extracted from sensor inputs (e.g., "I am 90% certain I saw the robotics book on the table"). We have developed an architecture that combines the complementary strengths of non-monotonic logical reasoning and probabilistic reasoning  to represent and reason with such descriptions. We can now pose interesting questions such as:

1. How to enable a robot to incrementally and interactively acquire domain knowledge from sensor inputs and high-level human feedback?

2. How to enable a robot to adapt its representation and reasoning at different resolutions to provide human-understandable explanations of beliefs, experiences and decisions?

Students will develop algorithms to address these questions in specific application domains, and evaluate these algorithms in simulation and on physical robot platforms.

Requirements:

• (Essential) Proficiency in probability theory, statistics, linear algebra and calculus.
• (Essential) Proficiency in an object-oriented language or scripted programming language of your choice.
• (Optional) Prior knowledge of logic or machine learning, and their use in robotics.
• (Optional) Prior experience of working with physical robots.

Investigator:

Dr. Mohan Sridharan

Theory of Intentions and Affordances for Assistive Robots

Mobile robots are increasingly being used to assist humans in complex domains characterized by partial observability and non-deterministic action outcomes. To truly collaborate with humans in such domains, a robot has to anticipate the intentions of its human collaborator, estimate the action capabilities (also known as affordances) of the collaborator, and offer assistance when the human's capabilities are likely to be insufficient to achieve the desired intention. We have developed theoretical claims about how to represent, reason with and learn these intentions and affordances. We can now pose interesting questions such as:

1. How to instantiate these thereotical claims in software architectures in the context of a robot collaborating with humans in complex domains?

2. How to exploit these theoretical claims to support reliable, computationally efficient, and explainable decision making?

Students will develop algorithms to address these questions in specific application domains, and evaluate these algorithms in simulation and on physical robot platforms.

Requirements:

• (Essential) Proficiency in probability theory, statistics, linear algebra and calculus.
• (Essential) Proficiency in an object-oriented language or scripted programming language of your choice.
• (Optional) Prior knowledge of logic or machine learning, and their use in robotics.
• (Optional) Prior experience of working with physical robots.

Investigator:

Dr. Mohan Sridharan