In order to endow robots with human-like abilities to characterize and identify objects, they must be provided with tactile sensors and intelligent algorithms to select, control, and interpret data from useful exploratory movements. Humans make informed decisions on the sequence of exploratory movements that would yield the most information for the task, depending on what the object may be and prior knowledge of what to expect from possible exploratory movements. This study is focused on texture discrimination, a subset of a much larger group of exploratory movements and percepts that humans use to discriminate, characterize, and identify objects. Using a testbed equipped with a biologically inspired tactile sensor (the BioTac), we produced sliding movements similar to those that humans make when exploring textures. Measurement of tactile vibrations and reaction forces when exploring textures were used to extract measures of textural properties inspired from psychophysical literature (traction, roughness, and fineness). Different combinations of normal force and velocity were identified to be useful for each of these three properties. A total of 117 textures were explored with these three movements to create a database of prior experience to use for identifying these same textures in future encounters. When exploring a texture, the discrimination algorithm adaptively selects the optimal movement to make and property to measure based on previous experience to differentiate the texture from a set of plausible candidates, a process we call Bayesian exploration. Performance of 99.6% in correctly discriminating pairs of similar textures was found to exceed human capabilities. Absolute classification from the entire set of 117 textures generally required a small number of well-chosen exploratory movements (median = 5) and yielded a 95.4% success rate. The method of Bayesian exploration developed and tested in this paper may generalize well to other cognitive problems.

No abstract

Successful performance of a sensorimotor task arises from the interaction of descending commands from the brain with the intrinsic properties of the lower levels of the sensorimotor system, including the dynamic mechanical properties of muscle, the natural coordinates of somatosensory receptors, the interneuronal circuitry of the spinal cord, and computational noise in these elements. Engineering models of biological motor control often oversimplify or even ignore these lower levels because they appear to complicate an already difficult problem. We modeled three highly simplified control systems that reflect the essential attributes of the lower levels in three tasks: acquiring a target in the face of random torque-pulse perturbations, optimizing fusimotor gain for the same perturbations, and minimizing postural error versus energy consumption during low- versus high-frequency perturbations. The emergent properties of the lower levels maintained stability in the face of feedback delays, resolved redundancy in over-complete systems, and helped to estimate loads and respond to perturbations. We suggest a general hierarchical approach to modeling sensorimotor systems, which better reflects the real control problem faced by the brain, as a first step toward identifying the actual neurocomputational steps and their anatomical partitioning in the brain.

No abstract

The notion that biological systems come to embody optimal solutions seems consistent with the competitive drive of evolution. It has been used to interpret many examples of sensorimotor behavior. It is attractive from the viewpoint of control engineers because it solves the redundancy problem by identifying the one optimal motor strategy out of many similarly acceptable possibilities. This perspective examines whether there is sufficient basis to apply the formal engineering tools of optimal control to a reductionist understanding of biological systems. For an experimental biologist, this translates into whether the theory of optimal control generates nontrivial and testable hypotheses that accurately predict novel phenomena, ideally at deeper levels of structure than the observable behavior. The methodology of optimal control is applicable when there is (i) a single, known cost function to be optimized, (ii) an invertible model of the plant, and (iii) simple noise interfering with optimal performance. None of these is likely to be true for biological organisms. Furthermore, their motivation is usually good-enough rather than globally optimal behavior. Even then, the performance of a biological organism is often much farther from optimal than the physical limits of its hardware because the brain is continuously testing the acceptable limits of performance as well as just performing the task. This perspective considers an alternative strategy called "good-enough" control, in which the organism uses trial-and-error learning to acquire a repertoire of sensorimotor behaviors that are known to be useful, but not necessarily optimal. This leads to a diversity of solutions that tends to confer robustness on the individual organism and its evolution. It is also more consistent with the capabilities of higher sensorimotor structures, such as cerebral cortex, which seems to be designed to classify and recall complex sets of information, thereby allowing the organism to learn from experience, rather than to compute new strategies online. Optimal control has been a useful metaphor for understanding some superficial aspects of motor psychophysics. Reductionists who want to understand the underlying neural mechanisms need to move on.