We characterized how hemiparetic stroke survivors and neurologically intact individuals adapt reaching movements to compensate for unpredictable environmental perturbations. We tested the hypotheses that like unimpaired subjects, hemiparetic stroke survivors adapt using sensory information obtained during only the most recent movements and that the reliability of target acquisition decreases as the degree of sensorimotor impairment increases. Subjects held the handle of a 2-joint robotic arm which applied forces to the hand while reaching between targets in a horizontal plane. The robot simulated a dynamic environment which varied randomly in strength from one trial to the next. The trial sequence of perturbations had a non-zero mean value corresponding to information about the environment which subjects might learn. Stroke subjects were less effective at adapting reaches to the perturbations than control subjects. From a family of potential adaptation models, we found that the compensatory strategy patients used was the same as that used by neurologically-intact subjects. However, analysis of model coefficients found that the relative weighting of prior perturbations and prior movement errors on subsequent reach attempts was significantly depressed post-stroke. Regulation of final hand position was also impaired in the paretic limbs. Measures of trajectory adaptation and final position regulation deficits were significantly dependent on the integrity of limb proprioception and the amount of time post-stroke. But whereas model coefficients varied systematically with impairment level post-stroke, variability of final positioning in the contralesional limb did not. This difference suggests that these two aspects of limb control may be differentially impaired post-stroke.