Robot servants who neat your house or cook your meal have long been the dream of science fiction writers and artificial intelligence researchers alike.
But according to researchers at the Computer Science and Artificial Intelligence Laboratory (CSAIL), if robots are ever to successfully navigate the ever-changing homes and workplaces to perform such complicated tasks, they will need to be more aware of their own limitations.
Today, the most effective robots are typically used either in fixed, carefully controlled environments such as manufacturing plants, or to perform fairly straightforward tasks such as vacuuming rooms, says Leslie Pack Kaelbling, a Panasonic professor of computer science and engineering at MIT.
Performing complicated sequences of actions in a cluttered, active environment such as a home will require robots to be more aware of what they don’t know, and therefore will need to learn it, Kaelbling says. This is because the robot cannot simply look around the kitchen and determine, for example, where all the containers are kept or what you would prefer to eat for dinner. To find out, you need to open the cabinets and look inside or ask a question.
“I would like to create a robot that could walk into your kitchen for the first time, having already been in other kitchens but not yours, and put away the groceries,” Kaelbling says.
In an article recently accepted for publication in the journal International Journal of Robotics Researchshe and CSAIL colleague Tomas Lozano-Perez describe a system designed for exactly this purpose, by continuously calculating the robot’s level of uncertainty about a given task, such as the location of an object or its own location in a room.
The uncertainty principle
The system relies on a module called the state estimation component, which calculates the probability that any object is where the robot thinks it is. That way, if the robot isn’t confident enough that an object is the one it’s 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.
For example, if a robot was trying to pick up a box of cereal off a shelf, it might decide that its uncertainty about the object’s position was too great to attempt to grab it. Instead, it would first take a closer look at the object to better understand its exact location, Kaelbling says. “He is always thinking about his beliefs about the world and how to change his beliefs 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 a given task by creating a plan for it in stages, using what the team calls hierarchical planning in the present.
“There’s this idea in AI that we’re very concerned about having the optimal plan, so we’re going to compute very hard over a long period of time to make sure we have a complete strategy formulated before we start executing,” Kaelbling says.
But in many cases, especially if the robot has a novel environment, it may not have enough knowledge of the area to make such a detailed plan in advance, he says.
Baby steps
Instead, the system creates a plan for the first stage of its task and begins implementing it before coming up with a strategy for the rest of the exercise. This means that instead of one enormous, complicated strategy that consumes a significant amount of processing power and time, the robot can create many smaller plans over time.
The downside to this process is that it can lead the robot to make stupid mistakes, such as picking up a plate and moving it to the table without realizing that it first needs to clear some space to put it down, Kaelbling says.
But such minor errors may be a price worth paying for more capable robots, he says: “When we try to get robots to do bigger and more complex things in more variable environments, we’re going to have to settle for some degree of suboptimality.”
In addition to home robots, the system could also be used to build more versatile industrial equipment or for disaster relief, Kaelbling says.
Ronald Parr, a professor of computer science at Duke University, says that most existing work on robot planning is typically divided into different groups working on specific, specialized problems. In contrast, Kaelbling and Lozano-Perez’s work breaks down the walls that exist between these subgroups and uses hierarchical planning to address the computational challenges that arise when trying to develop a more general-purpose problem-solving system. “Furthermore, it was demonstrated on a practical, general-purpose robotics platform that could be used for household or factory work,” Parr says.