Improving robots' planning abilities ‘one step at a time’
By Darren Quick
May 26, 2011
Although robots have started to creep into many homes in the form of robotic vacuum cleaners like the Roomba, many of us were hoping humanoid robots would be doing everything from preparing meals to doing the laundry by now. Unfortunately, it turns out that even tasks that we consider relatively simple can prove difficult for robots to replicate. Instead of attempting to program robots to devise a complete detailed plan before tacking a task, MIT scientists are programming them to apply the old adage of "one step at a time" so the robots break tasks down into smaller, easier to handle chunks.
Programming to provide robots with autonomous functions has traditionally been split into two complementary parts - task planning and geometric motion planning. Task planners can use partial descriptions to decide what needs to be done to accomplish a task. For example, when a robot is assigned a task to make a phone call, a task planner can decide that a room needs to be traversed to reach the phone, but not how to negotiate the furnishings to accomplish this task. Geometric motion planners on the other hand, can plan how to cross the room to the phone, but not what to do when it gets there.
The key to the approach taken by MIT scientists Leslie Kaelbling and Tomás Lozano-Pérez is how these two types of planning are integrated. Instead of programming a robot to construct a complete plan before tackling a task, they have devised a progressive algorithm that lets robots break a larger task down into smaller steps so it can concentrate on devising a plan for the first few and worry about the specifics of subsequent steps later.
As Lozano-Pérez tells MIT News, "we're introducing a hierarchy and being aggressive about breaking things up into manageable chunks."
The same approach is also applicable to geometric planning, with the robot only needing to construct a rough map of the area relevant to the task at hand, rather than expending much greater computational power on constructing a map of all the areas that the task will cover before it gets underway.
The researchers admit that devising a complete plan in advance is the best way to realize a goal in the fewest number of steps and ensure the plan can actually be achieved before it is put into action. However, they say their approach allows for much more flexibility as they are better able to deal with environments that may change over time.
In this respect, Kaelbling and Lozano-Pérez are going against robotics' traditional emphasis on optimizing behavior in favor of practicality.
"We're very consciously saying, 'No, if you insist on optimality then it's never going to be practical for real machines,'" says Lozano-Pérez.
With one of the keys to the approach revolving around deciding which decisions need to be tackled in advance and which ones can be left to worry about later, the researchers plan to integrate learning algorithms so the robots are better able to make such judgements.
As well as helping around the house, the researchers say their approach could also prove useful in logistics, military and surveillance applications.
Earlier this month, the MIT researchers presented a paper detailing their approach titled "Hierachical Task and Motion Planning in the Now," at the IEEE Conference on Robotics and Automation in Shanghai.
Source: MIT News