- Integrating Reinforcement-Learning, Accumulator Models, and Motor-Primitives to Study Action Selection and Reaching in Monkeys 2007-03-15
Ognibene, D. and Mannella, F. and Pezzulo, G. and Baldassarre, G.
Proceedings of ICCM 2006
- Metric State Space Reinforcement Learning for a Vision-Capable Mobile Robot 2007-04-17
V. Zhumatiy, F. Gomez, M. Hutter, and J. Schmidhuber: Metric State Space
Reinforcement Learning for a Vision-Capable Mobile Robot, Proc.
of the Int'l Conf. on Intelligent Autonomous Systems,
IAS-06, Tokyo, 2006.
- Toward an integrated biomimetic model of reaching. 2008-05-15
Caligiore D., Parisi D., Baldassarre G. (2007). Towards an integrated biomimetic model of reaching. Demiris Y., Scassellati B., Mareschal D. (eds.). The 6th IEEE International Conference on Development and Learning (ICDL2007). IEEE Catalog Number: 07EX1740C, ISBN: 1-4244-1116-5, Library of Congress: 2007922394, pp. 241-246. London: Imperial College
- Shared Intention Revisited: The Limits of Egoism are Not the Limits of Individualism 2007-04-10
Tummolini, L. (submitted to Economics and Philosophy)
- Building Robots with Analogy-Based Anticipation 2007-05-09
Petkov, G., Naydenov, Ch., Grinberg, M., Kokinov(2006); Proceedings of the KI 2006, 29th German Conference on Artificial Intelligence, Bremen
- Knowing the unknown 2013-03-27
- Robot butlers that tidy your house or cook you a meal have long been the dream of science-fiction writers and artificial intelligence researchers alike.
But if robots are ever going to move effectively around our constantly changing homes or workspaces performing such complex tasks, they will need to be more aware of their own limitations, according to researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL).
Most successful robots today tend to be used either in fixed, carefully controlled environments, such as manufacturing plants, or for performing fairly simple tasks such as vacuuming a room, says Leslie Pack Kaelbling, the Panasonic Professor of Computer Science and Engineering at MIT.
Carrying out complicated sequences of actions in a cluttered, dynamic environment such as a home will require robots to be more aware of what they do not know, and therefore need to find out, Kaelbling says. That is because a robot cannot simply look around the kitchen and determine where all the containers are stored, for example, or what you would prefer to eat for dinner. To find these things out, it needs to open the cupboards and look inside, or ask a question.
“I would like to make a robot that could go into your kitchen for the first time, having been in other kitchens before but not yours, and put the groceries away,” Kaelbling says.
And in a paper recently accepted for publication in the International Journal of Robotics Research, she and CSAIL colleague Tomas Lozano-Perez describe a system designed to do just that, by constantly calculating the robot’s level of uncertainty about a given task, such as the whereabouts of an object, or its own location within the room.
The system is based on a module called the state estimation component, which calculates the probability of any given object being what or where the robot thinks it is. In this way, if the robot is not sufficiently certain that an object is the one it is looking for, because the probability of it being that object is too low, it knows it needs to gather more information before taking any action, Kaelbling says.
So, for example, if the robot were trying to pick up a box of cereal from a shelf, it might decide its uncertainty about the position of the object was too high to attempt grasping it. Instead, it would first take a closer look at the object, in order to get a better idea of its exact location, Kaelbling says. “It’s thinking always about its own belief about the world, and how to change its belief, by taking actions that will either gather more information or change the state of the world.”
The system also simplifies the process of developing a strategy for performing a given task by making up its plan in stages as it goes along, using what the team calls hierarchical planning in the now.
“There is this idea in AI that we’re very worried about having an optimal plan, so we’re going to compute very hard for a long time, to ensure we have a complete strategy formulated before we begin execution,” Kaelbling says.
But in many cases, particularly if the environment is new to the robot, it cannot know enough about the area to make such a detailed plan in advance, she says.
So instead the system makes a plan for the first stage of its task and begins executing this before it has come up with a strategy for the rest of the exercise. That means that instead of one big complicated strategy, which consumes a considerable amount of computing power and time, the robot can make many smaller plans as it goes along.
The drawback to this process is that it can lead the robot into making silly mistakes, such as picking up a plate and moving it over to the table without realizing that it first needs to clear some room to put it down, Kaelbling says.
But such small mistakes may be a price worth paying for more capable robots, she says: “As we try to get robots to do bigger and more complicated things in more variable environments, we will have to settle for some amount of suboptimality.”
In addition to household robots, the system could also be used to build more flexible industrial devices, or in disaster relief, Kaelbling says.
Ronald Parr, an associate professor of computer science at Duke University, says much existing work on robot planning tends to be fragmented into different groups working on particular, specialized problems. In contrast, the work of Kaelbling and Lozano-Perez breaks down the walls that exist between these subgroups, and uses hierarchical planning to address the computational challenges that arise when attempting to develop a more general-purpose, problem-solving system. “What’s more, it is demonstrated on a practical, general-purpose robotic platform that could be used for domestic or factory work,” Parr says.
- A reinforcement-learning model of top-down attention based on a potential-action map 2008-05-15
Ognibene, D.; Balkenius, C. & Baldassarre, G.. 2008. A reinforcement-learning model of top-down attention based on a potential-action map. Pezzulo Giovanni, Rino Falcone, Castelfranchi Cristiano (ed.). The Anticipatory Approach. Berlin. Springer-Verlag.
- Mapping neurological disease 2012-09-05
- Disorders such as schizophrenia can originate in certain regions of the brain and then spread out to affect connected areas. Identifying these regions of the brain, and how they affect the other areas they communicate with, would allow drug companies to develop better treatments and could ultimately help doctors make a diagnosis. But interpreting the vast amounts of data produced by brain scans to identify these connecting regions has so far proved impossible.
Now, researchers in the Computer Science and Artificial Intelligence Laboratory at MIT have developed an algorithm that can analyze information from medical images to identify diseased areas of the brain and their connections with other regions.
The MIT researchers will present the work next month at the International Conference on Medical Image Computing and Computer Assisted Intervention in Nice, France.
The algorithm, developed by Polina Golland, an associate professor of computer science, and graduate student Archana Venkataraman, extracts information from two different types of magnetic resonance imaging (MRI) scans. The first, called diffusion MRI, looks at how water diffuses along the white-matter fibers in the brain, providing insight into how closely different areas are connected to one another. The second, known as functional MRI, probes how different parts of the brain activate when they perform particular tasks, and so can reveal when two areas are active at the same time and are therefore connected.
These two scans alone can produce huge amounts of data on the network of connections in the brain, Golland says. “It’s quite hard for a person looking at all of that data to integrate it into a model of what is going on, because we’re not good at processing lots of numbers.”
So the algorithm first compares all the data from the brain scans of healthy people with those of patients with a particular disease, to identify differences in the connections between the two groups that indicate disruptions caused by the disorder.
However, this step alone is not enough, since much of our understanding of what goes on in the brain concerns the individual regions themselves, rather than the connections between them, making it difficult to integrate this information with existing medical knowledge.
So the algorithm then analyzes this network of connections to create a map of the areas of the brain most affected by the disease. “It is based on the assumption that with any disease you get a small subset of regions that are affected, which then affect their neighbors through this connectivity change,” Golland says. “So our methods extract from the data this set of regions that can explain the disruption of connectivity that we see.”
It does this by hypothesizing, based on an overall map of the connections between each of the regions in the brain, what disruptions in signaling it would expect to see if a particular region were affected. In this way, when the algorithm detects any disruption in connectivity in a particular scan, it knows which regions must have been affected by the disease to create such an impact. “It basically finds the subset of regions that best explains the observed changes in connectivity between the normal control scan and the patient scan,” Golland says.
When the team used the algorithm to compare the brain scans of patients with schizophrenia to those of healthy people, they were able to identify three regions of the brain — the right posterior cingulate and the right and left superior temporal gyri — that are most affected by the disease.
In the long term, this could help drug companies develop more effective treatments for the disease that specifically target these regions of the brain, Golland says. In the meantime, by revealing all the different parts of the brain that are affected by a particular disorder, it can help doctors to make sense of how the disease evolves, and why it produces certain symptoms.
Ultimately, the method could also be used to help doctors diagnose patients whose symptoms could represent a number of different disorders, Golland says. By analyzing the patient’s brain scan to pinpoint which regions are affected, it could identify which disorder would create this particular disruption, she says.
In addition to schizophrenia, the researchers, who developed the algorithm alongside Marek Kubicki, associate director of the Psychiatry Neuroimaging Laboratory at Harvard Medical School, are also investigating the possibility of using the method to study Huntington’s disease.
Gregory Brown, associate director of clinical neuroscience at the University of California at San Diego’s Center for Functional MRI, who was not involved in developing the model, plans to use it to study the effects of HIV and drug addiction. “We will use the method to gain a clearer perspective on how HIV infection and methamphetamine dependence disrupts large-scale brain circuitry,” he says.
The method is a critical step away from studying the brain as a collection of localized regions toward a more realistic systems perspective, he says. This should assist the study of disorders such as schizophrenia, neurocognitive impairment and dementia associated with AIDS, and multiple sclerosis, which are best characterized as diseases of brain systems, he says.