One of the new abilities that Berenson and his group will help robots attain is navigating unfamiliar environments by reusing previous experience from similar environments. When the robot sees debris-covered stairs, for example, it might skip straight to the plan that worked last time it saw stairs, then adapt to the new details. The key, Berenson says, will be in teaching robots to better adapt to unknowns.
“Current robots don’t understand how to interact with the physical world very well,” said Berenson. “The approaches we have now are very conservative when it comes to being able to work with uncertainty.”
On a swaying ship, even the floor beneath a robot’s feet is in motion. That’s why Berenson’s group is focusing on how robots can use their hands to handle unexpected forces, by bracing against walls or using other objects as supports.
Berenson works closely with a team at Virginia Tech, which built and operates the SAFFiR humanoid robot. SAFFiR is designed to help fight fires on ships, but to be truly useful in an emergency, the robot needs to be able to handle tight corridors and steep slopes that don’t look and feel like they normally do – perhaps the floor is warped from heat or covered in debris from an explosion.
Through computer simulations, Berenson and his students test out SAFFiR’s ability to walk over difficult terrain. They design algorithms that build rough maps of the floor and walls with SAFFiR’s laser scanner and make quick decisions about foot and hand placements. Conventional robots would scan the whole area and consider many possible routes before beginning to move, but this is too slow for good cooperation with humans – let alone in an emergency.
The grant is also intended to help robots graduate from handling rigid objects to handling softer materials that can change shape. Ropes, chains and o-rings can tangle or catch on protrusions.
“How do we make the execution of these plans robust? It’s not going to be by making a plan and sticking to it no matter what,” said Berenson. “We’re going to make a policy with contingencies.”
For instance, a fire-fighting robot like SAFFiR would need to understand how a hose moves and have the capacity to troubleshoot kinks and snags.
Berenson and his team are currently setting up a pair of robotic arms and hands to develop a set of algorithms that can observe a demonstration with a previously unfamiliar object and then explore the forces that object can handle. As with navigation, the robot would build up a library of objects and specifications that could then be used as a starting point for manipulating new materials.
This work is supported by the Office of Naval Research, grant number N00014-17-1-2050.
by Katherine McAlpine
December 1, 2016