Using the finger-to-thumb opposition sequence (FOS) learning task, we characterized motor skill learning in sub-acute patients hospitalized for rehabilitation following traumatic brain injury (TBI). Ten patients (Trained TBI) and 11 healthy participants (Trained Healthy) were trained using a multi-session protocol: a single session was afforded in the first week of the study, and four daily sessions were afforded during the second week. Intensity of practice was adapted to patients. Performance speed and accuracy were tested before and after each session. Retention was tested 1 month later. Ten patients (Control TBI) had no FOS training and were tested only at the beginning and the end of the 6 week period. Although baseline performance on the FOS was very slow, all three phases of skill learning found in healthy adults (acquisition, between-session consolidation gains, and long-term retention) could be identified in patients with TBI. However, their time-course of learning was atypical. The Trained TBI group improved in speed about double the spontaneous improvements observed in the Control TBI group, with no speed-accuracy tradeoff. Normalized to their initial performance on the FOS, the gains accrued by the Trained TBI group after a first training were comparable to those accrued by healthy adults. Only during the second week with daily training, the rate of improvement of the Trained TBI group lagged behind that of the Trained Healthy group, due to increasing within-sessions losses in performance speed; no such losses were found in healthy participants. The Functional Independence Measure scores at the start of the study correlated with the total gains attained at the end of the study; no correlations were found with severity of injury or explicit memory impairments. Despite within-sessions losses in performance, which we propose reflect cognitive fatigue, training resulted in robust overall learning and long-term retention in patients with moderate-severe TBI. Given that the gains in performance evolved mainly between sessions, as delayed, offline, gains, our results suggest that memory consolidation processes can be effectively engaged in patients with TBI. However, practice protocols and schedules may need to be optimized to better engage the potential for long-term plasticity in these patients.