A puzzling problem: Teaching robots to criticize themselves

Self-assessment is a huge roadblock on the road to versatile automation
Written by Greg Nichols, Contributing Writer

The flubbed comeback, the missed opportunity, the flagrant mistake -- it can seem like we're hard-wired to dwell on the ways we come up short.

Now machines may be getting that capacity.

Google Duplex beat the Turing test: Are we doomed?

As part of the competitive Multi-Disciplinary University Research Initiatives (MURI) program, the Department of Defense just awarded $7.5 million to researchers from Carnegie Mellon, Brigham Young, UMass Lowell, and Tufts to develop new methods for machine self-assessment.

The research lies at the intersection of machine learning and artificial intelligence, and it comprises what's likely to be one of the hottest areas of robotics development over the next decade.


The reason is easily grasped: Robots are becoming increasingly autonomous, and engineers can't possibly account for every decision an autonomous system will make. That becomes increasingly true as robots begin operating outside structured environments like factories and warehouses.

We humans have a nifty trick in the form of a constant feedback loop, which reinforces behaviors that lead to positive results, enables us to transfer those behaviors to similar situations, and, ideally, helps us avoid decisions that lead to negative results.

But while that capacity is simple enough to describe, it's devilishly tricky to transplant into a machine.

A big reason is the fundamental complexity of assigning value to disparate outcomes. Though far from perfect, humans are actually pretty good at that. Understanding what it means to be a bad gardener, for example (i.e., killing plants), allows me to draw conclusions about the behaviors that make me a bad gardener (failure to water).

Those realizations, which stem from value judgments about good and bad gardening, help me approach related but completely new situations, such as caring for a pet.

Programming a robot to gauge its performance at a discreet and repeatable task is straightforward. Is the screw placed correctly? Is the joint welded? But giving a robot the tools to intuit how a good outcome in either of those situations can lead to a good outcome in a situation it has yet to encounter is something else entirely, one that requires a sophisticated transfer of value judgements and of lessons learned.

"You'd like the robot to be able to explain why it can or why it can't do a task," explains Aaron Steinfeld, a professor at the Robotics Institute at Carnegie Mellon University.

The MURI grant will run for five years, during which researchers will primarily work with robots tasked with manipulating objects by feel and transferring lessons to new situations.

The efforts could lead to robots that can do all sorts of on-the-fly work, such as urban search and rescue and emergency repair.

Related and previous Coverage

Robots with soft hands will transform the world. Here's why.

Labor-starved markets are waiting for a new class of technology that brings dexterity and nimbleness to industrial robots

'Inspector Gadget' arm makes drones more useful (and a little scary)

One of the big stumbling blocks for commercial drones is loading and unloading payload. This lightweight origami arm offers a novel solution.

"First-ever" use of a robot in a military breaching exercise

U.S. and allies are escalating the use of robots in training exercises, increasing speculation that unmanned platforms will see heavier use on tomorrow's battlefields.

Ransomware for robots is the next big security nightmare

Researchers found they were able to infect robots with ransomware; in the real world, such attacks could be highly damaging to businesses if robotic security isn't addressed.

Editorial standards