Starting Position Target Position Visual Feedback Actual Hand Position 15 cm B D C A Feedback Uncertainty Small Blur Large Blur

models make specific predictions about the role of feedback and state estimation uncertainty in estimating movement-related vari-For example, when playing tennis we might observe that we return the ball short of an intended location. We will then adapt our next swing to this error. To optimally perform the task, we will adapt more if we clearly see the error as compared to if we only see it through peripheral vision (high feedback uncertainty) and that we will adapt more if we are still warming up for today's practice than if we have been playing for a while (low state estimation uncertainty). The Bayesian predictions for trial-by-trial adaptation are rather intuitive: across-trial adjustment is more pronounced when state estimates are more uncertain and/or when estimates in feedback are more certain to the nervous system. The present study tests these predictions using a trial-by-trial adaptation task with visual perturbations. We compare the Bayesian model with state space models (e.g., Cheng and Sabes, 2006), both of which are frequently used for studying motor learning. We also compare to models that combine features of both state space models and Kalman models. In Experiment 1 we manipulate feedback uncertainty by blurring the visual feedback. In Experiment 2 we provide feedback of different quality for fixed periods of time to leave the subject with defined state estimation uncertainty. As qualitatively predicted by Bayesian models, adaptation is significantly faster with less feedback uncertainty and with more state estimation uncertainty. These results suggest that the nervous system represents feedback uncertainty and state estimation uncertainty and uses knowledge of uncertainty during motor adaptation.