OBJECTIVES: To determine whether training influenced the accuracy of observational kinematic assessment (OKA) of hemiplegic upper extremity impairment and to elucidate the contribution of knowledge of results to learning. DESIGN: Intervention study; before-after testing of OKA accuracy after training, using 2 trained groups (with knowledge of results, KR group; without, NKR group) and 1 control group with a 1-week retention test. SETTING: Tertiary teaching. PARTICIPANTS: Fifty-one first-year physical therapy student volunteers acted as observers. They were ranked on pretest accuracy and then randomized into groups. A consecutive sample of 11 stroke patients and 4 nondisabled subjects acted as performers. INTERVENTIONS: Performers were videotaped with 3 cameras and upper extremity kinematics derived using computer-assisted motion analysis. Training and test videotapes were generated. Training groups received video-based training of path indirectness accuracy on 4 occasions. The OKA accuracy of all observers' judgments of speed, jerkiness, and path indirectness were examined pretest and posttest. MAIN OUTCOME MEASURE: Accuracy reported as mean absolute error, which was calculated as difference between observers' judgments and criterion kinematic values. RESULTS: The KR and NKR groups showed reduction in mean absolute error after training of 34.8% and 6.2%, respectively. Improvements were retained after a 1-week no intervention period. Transfer to trained, but not untrained kinematic parameters occurred. The control group did not change. CONCLUSIONS: OKA accuracy is susceptible to training and knowledge of results aids learning. However, training is task specific.