Though robots are becoming more sophisticated, they are still unable to ask questions. Robots can request data – “please say or enter your order number” – but that’s not the same as knowing when and how to ask questions.
For a human dealing with an automated customer service robot over the telephone or online chat box, that might be an irritation. But in the middle of a firefight, a robot tank that can’t ask for clarification on whether that unidentified object is an enemy vehicle might be a problem.
However, U.S. Army scientists have developed algorithms that give the robots the ability to systematically ask questions.
“We created a computational model for automated question generation and learning,” said Felix Gervits, a researcher at the Combat Capabilities Development Command at the Army Research Laboratory (ARL). “The model enables a robot to ask effective clarification questions based on its knowledge of the environment and to learn from the responses. This process of learning through dialogue works for learning new words, concepts and even actions,” he added.
The project, which also included researchers from Tufts University and the Naval Research Laboratory, is based on the Human-Robot Dialogue Learning system, which enables machines to learn by asking questions.
“Researchers present an expanded dialogue example using the computational model for question generation in which a robot is verbally instructed to perform a task in an unexplored environment,” according to an ARL news release. “The robot is able to learn the names of objects and locations as well as goals by asking clarifying questions.”
The inability to properly clarify isn’t a problem limited to robots. Many military mishaps – such as collisions and friendly fire incidents – have occurred because human conversation frequently involves imprecise questions or answers. The key is to get robots to ask questions that elicit precise answers from humans, especially during noisy or stressful situations, when people may have difficulty issuing clear voice commands.
Human says Army robot does
ARL presented a sample dialogue between a human, and a robot using the new algorithms to methodically query a command to move an optimized synthesizer instrument.
Human: “Pick up the optimized synthesizer.”
Robot: “What does the optimized synthesizer look like?”
Human: “Synthesizers are measurement tools with a number display and a dial. There are three of them.”
Robot: “Is it blue?”
Robot: “What do I do with it?”
Human: “Move to the primary cabinet.”
Robot: “Then what?”
Human: “Place it on shelf B.”
To a non-roboticist, this dialogue may seem child-like. But it is actually a logical sequence of queries and responses to narrow down a range of potential actions into a single correct choice, such as placing the right object on the right shelf. In many ways, this is how humans themselves learn.
For now, the Army indicates that the new system is useful for more routine tasks. “Such technology can be employed on robots in remote collaborative interaction tasks such as reconnaissance and search-and-rescue, or in co-located human-agent teams performing tasks such as transport and maintenance,” ARL noted.
However, it is easy to envision scenarios where this technology could be used in combat. An unmanned tank or drone encounters an unfamiliar object, such as a new enemy vehicle that U.S. forces have not encountered, and thus is missing from the robot’s database of enemy equipment. Or, an automated staff AI needs more data to help commanders devise battle plans, or a next-generation manned tank requires information to assist the crew with navigating the vehicle. In those cases, the ability to ask the right questions is vital.
Feature Image: Running Man robot of Team IHMC Robotics from Pensacola, FL, clears a doorway during Defense Advanced Research Projects Agency (DARPA) Robotics Challenge. (U.S. Navy photo by Greg Vojtko/Released)