Click on the image to the left to see a movie of Thing displaying some of its earliest walking gaits.

The first policy Thing learned was how to rotate in place. Here's a movie showing the training, the number in the lower left of the frame is the elapsed time of the learning episode. The approach we employed made it possible for Thing to learn this gait in about 11 minutes, on-line, in a single training episode. Knowledge of the motor synergies involved in rotating under varying conditions significantly improved the acquisition of other behavior. For instance, Thing learned to translated in roughly half the time given the prior rotate policy than it did without it, and ultimately, the average amount of translation per action was roughly double as well. We believe that this phenomenon is consistent with the kind of staged, sequential development that human neonates exhibit as observed by Developmental Psychologists like Piaget.

Policies built using this framework for perceiving state and executing control are very robust to changing conditions. The picture at right shows Thing lying on its back and manipulating a spherical globe using the rotate walking gait that was acquired on a flat terrain (sorry about the picture quality). Check out these videos of our four fingered Utah/MIT hand using the same policy to rotate a cylinder (soda can) and screw in a lightbulbs.

In the following demonstration, Thing knows its start position (leftmost X), a relative displacement to the goal (rightmost X), and has no knowledge of the intervening obstacle. He uses his forward-looking IR proximity detectors to observe obstacles enroute and maps them into his evolving configuration space (top row). Thing continuously replans paths to the goal using harmonic functions while updating the configuration space obstacle map as it acquires information about the world. The locomotion plan follows a streamline in harmonic function by selecting one of two temporally extended actions (ROTATE-GAIT, or TRANSLATE-GAIT) in a 4 state finite state automaton. The state is derived from a 2 bit "interaction-based" state descriptor. One bit describes the convergence status of the ROTATE-GAIT action, and the other describes the convergence status of the TRANSLATE-GAIT action with respect to the current path plan. The control knowledge represented in these two actions and this FSA are adequate for finding paths in any cluttered world provided that a path exists at the resolution of our configuration space map.

In about 20 minutes, Thing traveled approximately 5 meters and accrued a translational odometry error of approximately 5 cm. You can see an accelerated movie of this experiment by clicking here.