Background: Wearable ankle robotics could potentially facilitate intensive repetitive task-specific gait training on stair environment for stroke rehabilitation. A lightweight (0.5kg) and portable exoskeleton ankle robot was designed to facilitate over-ground and stair training either providing active assistance to move paretic ankle augmenting residual motor function (power-assisted ankle robot, PAAR), or passively support dropped foot by lock/release ankle joint for foot clearance in swing phase (swing-controlled ankle robot, SCAR).In this two-center randomized controlled trial, we hypothesized that robot-assisted gait training using either PAAR or SCAR in stair environment are more effective to enhance gait recovery and promote independency in early stroke, than conventional training.Methods: Sub-acute stroke survivors (within two months after stroke onset) received hospital usual care plus 20-session robot-assisted training (at least twice weekly, 30-minute per session) on over-ground and stair environments, wearing PAAR (n=14) or SCAR (n=16), as compared to control group receiving conventional training only (CT, n=17). PAAR provided 2.5Nm-4.2Nm torque which was calibratedin each training session to adjust for any progression of functional changes throughout the intervention. Clinical assessments were performed before and after the 20-session intervention, including functional ambulatory categoryas primary outcome measure, along with Berg balance scale and timed 10-metre walk test.Results: After the 20-session interventions, all three groups showed statistically significant and clinically meaningful within-group functional improvement in all outcome measures (p<0.005). Between-group comparison showed SCAR had greater improvement in functional ambulatory category(mean difference +0.6, medium effect size 0.610) with more than 56% independent walkers after training, as compared to only 29% for CT. Analysis of covariance results showed PAAR had greater improvement in walking speed than SCAR (mean difference +0.15m/s, large effect size 0.752), which was in line with the higher cadence and speed when wearing the robot during the 20-session robot-assisted training over-ground and on stairs.Conclusions: Robot-assisted stair training would lead to greater functional improvement in gait independency and walking speed than conventional training in usual care. The active powered ankle assistance might facilitate users to walk more and faster with their paretic leg during stair and over-ground walking.Trial Registration:ClinicalTrials.gov NCT03184259. Registered on 12 June 2017.