Movement triggered by sensory stimuli requires that the networks generating the motor commands receive an adequate driving input, which, in general, is a transformed version of the initial sensory signal. We investigated the nature of this transformation in a task in which monkeys categorize the speed of tactile stimuli as either low or high, reaching for one of two pushbuttons to indicate their choice. Extracellular recordings from primary motor cortex revealed two types of neurons selective for the speed categories: ones that fire at higher rates for low versus high speeds, and others that do the opposite. These differential responses are task-specific; no firing rate modulation was seen when identical arm movements were triggered by visual cues or when stimuli were delivered passively. Analyses using decoding and modeling techniques produced two main results. First, the neurons accurately encode the chosen category; an observer measuring their responses can exhibit a psychophysical performance during categorization identical to the monkey's. Second, by analyzing separately the trials in which hits and errors were scored, it is possible to distinguish purely sensory activity from activity exclusively related to arm motion. The recorded responses did not match either of these alternatives but were consistent with a model in which the category-tuned neurons are the link between the output of the sensory categorization process and the motor command used to indicate the animal's decision. Thus, the observed activity seems to encode a preprocessed version of the sensory stimulus and to participate in driving the arm motion.Key words: primary motor cortex; primary somatosensory cortex; sensorimotor transformations; categorization; tactile motion; decoding; modeling A large body of evidence has accumulated establishing that primary motor cortex (M1, area 4) is involved in the control of voluntary movements (Evarts, 1981;Georgopoulos, 1995). Neuronal activity in this area strongly correlates with the parameters of arm motion, such as force and direction (Georgopoulos et al., , 1992Schwartz et al., 1988;Johnson et al., 1996), and with the geometry and mechanics of the joints (Thach, 1978;Caminiti et al., 1990Caminiti et al., , 1991Werner et al., 1991;Scott and Kalaska, 1997). Activity in M1 related to sensory events or cues has been reported too (Lamarre et al., 1983;Martin and Ghez, 1985). Using paradigms that involve the manipulation of sensory information as well as the execution of arm movements, other studies have uncovered complex responses not uniquely related to motor performance but instead reflecting either sensory processing or intermediate sensorimotor representations Crutcher and Alexander, 1990;Hocherman and Wise, 1991;Mountcastle et al., 1992;Ashe et al., 1993;Riehle et al., 1994;Pellizzer et al., 1995;Shen and Alexander, 1997; Zhang et al., 1997). If well characterized, the stimulus-related signals at the sensorimotor interface should provide insight into the nature of the neural computations implement...