Towards Robots with Human Dexterity

 It is probably self-evident that humans vastly out-perform robots, especially in tasks requiring dexterity, contact, and physical interaction with the world (e.g. as required to use tools). It may be less obvious that human 'hardware' (e.g. muscles) and 'wetware' (e.g. nerves) are orders of magnitude slower than those of modern robots. How do we do it? How do we achieve our spectacular performance despite these profound limitations?

 The over-arching goal of this collaborative research project was to improve robot performance and facilitate seamless human-robot physical collaboration. The way we approached this goal was:

• first, to study how humans perform simple physical interaction tasks that robots typically find challenging;
• second, to test a quantitative mathematical theory of how humans accomplish their superior performance; and
• third, to demonstrate that a low-cost robot programmed according to this theory could successfully perform tasks requiring contact and physical interaction.

The tasks we studied included: managing kinematic constraints (as when opening a door); and managing complex dynamic interactions (as with liquid sloshing in a cup).

 The quantitative theory we tested is that humans accomplish tasks like these by using a repertoire of well-learned and stereotyped 'primitive' dynamic actions which can be evoked by the central nervous system. The essence of this theory is that these 'dynamic primitives' require minimal central intervention as they play out and produce behavior, thereby working around the slow response of the neuromechanical system.

Intellectual Merit

Our first key finding was that human performance of kinematically-constrained motions (like opening a door) is quite different from modern robot programming. Usually, robot programming tries to control motion in directions where minimal or no opposing force is expected; and control force in directions where motion is impossible (due to a constraint). Parsing these two sets of directions can be computationally challenging.

Instead, human performance takes advantage of neuro-muscular mechanical impedance (essentially, the compliant behavior presented by the limbs) to manage the physical interaction without requiring precise knowledge of the external constraint. Interestingly, while this is beneficial (a 'feature') because it mitigates the need for precise control of either force or motion, we showed that it is also a drawback (a 'bug') insofar as humans are unable to control force independent of motion.

Our second key finding was that robot programming based on these primitive dynamic actions is remarkably effective. We showed its ability to manage:

• control of a robot arm where the desired end-point (hand) behavior is not sufficient to define the behavior of the robot joints;
• coupled closed-chain operation of two arms (without solving the difficult algebraic problem of  finding joint motions compatible with contact between the two arms);
• seamless operation into and out of singular configurations (where conventional motion-control approaches lack a well-defined mathematical solution);
• seamless transition into and out of contact (without switching between motion control and force control);
• guaranteed stable interaction with a poorly-modeled object; and
• rapid execution of motions while in contact.

Broader Impact

Taken together, our results provide new insight about how the human neuro-muscular system achieves its remarkable performance, but also insight about concomitant limitations--what we are not good at as well as what we do well. This deeper knowledge of human performance limitations as well as capabilities promises to enable better approaches to robot-aided rehabilitation of persons recovering from injury, including neurological injury such as cerebro-vascular accident (stroke).

Our results also support a new approach to robot programming and control based on composition of primitive dynamic actions. A patent describing this programming approach issued in 2019. It is especially important for robots intended to contact and physically collaborate with humans. Robots that exhibit behavior composed in a way similar to human movement will be fundamentally more predictable by humans. We believe that they will also be more comprehensible to humans, and hence fundamentally more trustworthy and acceptable to humans. This will facilitate future seamless integration of robotic technology into daily life and work. The societal impact (both positive and negative) could be profound.

Last Modified: 03/16/2021
Modified by: Neville Hogan