This 'Psychic Robot' Is Here to Fix Your Mistakes Before They Happen

Hey, we're only human, right? We have human brains with human reflexes and human memories. We have human accidents and suffer human consequences—often fatally, it turns out. We are but biological machines, slapped together over aeons by way of advantageous accidents.

But maybe we can fix this particular aspect of human-ness, or at least start to. A group of bioengineers at the University of Illinois at Chicago has come up with an algorithm capable of divining human intentions from physical observations, even if the action being analyzed is interrupted before being completed.

Say a driver hits a patch of ice going into a curve and loses control; the algorithm would be able to redirect the car back onto its intended course. This is all described in an open-access paper out now in the journal PLOS One: "I Meant to Do That: Determining the Intentions of Action in the Face of Disturbances."

"Disturbances, distractions, and pathologies can interfere in many situations and prevent our actions from matching our intent," write Justin Horowitz and James Patton, co-authors of the new paper. "A pilot, for example, may fail to complete a maneuver because of turbulence. Such challenges are broadly present in many human-machine interactions. One may speculate on how we might use these very same machines to elucidate the underlying intention. Such a possibility would be broadly useful in any area where intended actions might be thwarted by disturbances."

Specifically, the researchers tested a scenario in which a person's hand (either simulated or an actual human hand) goes to reach something on the other side of a desk. Mid-reach, the arm is batted away, leaving the algorithm to quickly determine what it had intended to be doing and then take over, completing the action. Or computing the completed action, at least.

Horowitz and Patton call their method "intent extraction." It consists of a set of filters that are able to infer the arm's intended trajectory based on the nature of the interruption, the environmental dynamics encountered by the arm, the biomechanics of the arm itself, and the arm's actual motion at the time of its interruption. All of this is then jammed together in a gnarly looking differential equation and solved, with the result being the likely trajectory had the system not been interrupted—or, in other words, the equilibrium state of all of those competing influences, from muscles to nervous system.

"The arm has a dynamic equilibrium in the sense that the mass of the arm resists unexpected acceleration and the elasticity of the arm's muscles resists unexpected stretching," Horowitz told me. "I say unexpected because the last term in [the equation] describes how even though mass resists all forces and springs resist all stretches, in the context of well-planned movement they only resist disturbances to the plan."

The pair first tested out their algorithm using the reaches of a simulated right arm, which were disturbed by two different types of forces: pulses and noise. Pulses are just concentrated forces directed at right angles to the arm's direction of motion, while noise forces are, much as they sound, chaotic multi-directional forces applied over a span of time. We can imagine a pulse as someone slapping the reaching hand away, and noise as the turbulent flow of air or water around the arm.

In a second part to the experiment, they followed up with human subjects, compiling data on 730 total reaches. Within 150 milliseconds following the disturbances, the algorithm could predict intended hand paths well enough. After that, around the point that the brain begins to react, things got murky.

"The algorithm does not predict the future," Horowitz explained. "It can only recover what the person is intending in real-time. On modern hardware, that means the algorithm is a fraction of a millisecond behind reality. That fraction of a millisecond is still many many times faster than the brain can react. A helpful analogy might be anti-lock brakes. A person uses constant pressure on the brakes and that means they want to stop, but pumping the brakes much faster than a human could manage is actually more effective. Similarly, the car's computer can react faster at the steering wheel and it can control individual wheels."

The brain catches up to reality after about 100 milliseconds, which may be too late already to sufficiently react. A computation isn't instantaneous, but it winds up being fast enough to matter, at least.

This has implications for not just cars and other large machines piloted by humans, but in medicine as well. An example offered by the researchers is that of a stroke patient beset by tremors. They may be able to intend to lift the cup to their mouth, but are foiled by the tremors. The algorithm could offer a solution.