Research

I'm currently working in the Stanford AI department under Prof. Jean-Claude Latombe. My work on motion planning for legged robots has been applied to the LEMUR free-climbing robot (JPL), the HRP-2 humanoid robot (AIST of Japan), the ATHLETE six-legged lunar robot (NASA/JPL), the LittleDog robot (Boston Dynamics Inc), the Capuchin free-climing robot (Stanford), and the Asimo robot (Honda).

Why legged robots?

Legged robots have a potential advantage over wheeled vehicles in navigating rough and steep terrain, because the robot can remain stable by carefully choosing where to place its feet. A vehicle that could cover rocky, steep, and forested terrain could have numerous applications in interplanetary exploration, military transportation, construction, and even recreational use. Humanoid robots have a potential advantage of being able to navigate and perform tasks in human-oriented environments with human-oriented tools.

Why aren't legged robots in common use?

Though many legged robots have been built, operating in complex environments poses several challenges. For example, building a 3D model of the terrain is difficult in areas with thick vegetation. For humanoid robots, the name, purpose, and appearance of household objects will need to be recognized, and they will need to translate human commands into appropriate actions. These problems are still being actively researched.

My research has primarily focused on motion planning -- coordinating the motions of the robot's joints in order to achieve a task. Motion planning is particularly important when it matters greatly where the robot places its feet or interacts with the environment -- one ill-placed step on a rocky slope could be the robot's last. This is precisely the situation in which legged robots have an advantage over wheeled vehicles.

What is the current status of your research?

I have built a motion planner for legged robot navigation and another for humanoid robot object manipulation.

The first planner can produce motions that traverse highly irregular terrain. For less irregular terrain, the planner can quickly plan by repeating high-quality example motions (motion primitives). It works with a huge variety of robot morphologies with little adjustment, and has been tested on the HRP-2, ATHLETE, LittleDog, and Capuchin robots in simulation and on the Capuchin in real experiments.

The second planner can produce motions for the ASIMO to manipulate objects on a table by pushing. It has been tested in simulation and experiments, and is able to produce motions in about a minute for most problems.

What are your short-term goals?

For legged locomotion, I hope to improve the speed of the planner for most terrain. It seems very likely that a planner for extremely difficult terrain will be slow, because even a human climbing over boulders can be puzzled by the number of alternative routes. But in most cases, a robot will be on gentle terrain, or a solution will be obvious. In these cases, the planner should act quickly. Furthermore, I hope to include map-building and model uncertainty into the planner. I would not be surprised if intelligent walking machines can be built in the next decade or two.

For manipulation, I hope to develop a full planning system for multi-handed, dexterous and non-dexterous object manipulation.

Projects

	The "contact-before motion" approach for legged locomotion planning Project page
	Improving planning using learning and motion primitives Project page
	Humanoid robot manipulation planning Project page

Old research and other projects