Spinal cord injury (SCI) affects approximately 300,000 individuals in the United States and 2.5 million worldwide, with motor-complete injuries resulting in permanent loss of ambulation. Robotic exoskeletons have emerged as promising rehabilitation tools, but current devices employ fixed or pre-programmed assistance patterns that fail to adapt to individual patient characteristics, injury level, and real-time performance. This paper presents the design and protocol for a randomized controlled trial (RCT) evaluating RL-Exo, a reinforcement learning (RL)-based personalized exoskeleton assistance system for gait training in individuals with chronic, motor-complete SCI (neurological levels T3-T12). The RL framework continuously learns patient-specific assistance parameters (hip and knee joint torque profiles, timing of stance-to-swing transition, lateral balance support) by optimizing a reward function that combines kinematic alignment with able-bodied gait templates, user-reported comfort, and energy expenditure (estimated via heart rate and oxygen consumption). The protocol randomizes 60 participants (target enrollment) into three arms: (1) RL-Exo with personalized assistance (N=24), (2) fixed-assistance exoskeleton with standard predefined gait pattern (N=24), and (3) conventional overground gait training without exoskeleton (N=12). The intervention period is 12 weeks (3 sessions per week, 60 minutes per session). Primary outcomes include the 6-Minute Walk Test (6MWT) distance and Walking Index for Spinal Cord Injury II (WISCI-II) assessed at baseline, 6 weeks, 12 weeks, and 6-month follow-up. Secondary outcomes include gait quality metrics (symmetry index, hip-knee coordination), user satisfaction (Quebec User Evaluation of Satisfaction with Assistive Technology, QUEST 2.0), energy cost of walking (oxygen consumption per meter), falls and adverse events, and RL algorithm metrics (reward convergence, adaptation rate). We hypothesize that RL-Exo will achieve superior improvements in walking distance (primary outcome: 6MWT improvement of ≥100 meters at 12 weeks vs. ≤40 meters for fixed-assistance, p<0.01) and user satisfaction, with fewer falls and lower energy cost. The RL algorithm learns optimal assistance patterns within the first 4-6 sessions (approximately 240-360 minutes of walking) and continues to micro-adapt thereafter. Personalized patterns differ significantly by injury level: T3-T6 (high thoracic) requires greater lateral balance support, T7-T10 (mid-thoracic) requires increased hip flexion assistance, and T11-T12 (low thoracic) requires primarily ankle-foot coordination. This protocol represents the first RCT of RL-based personalized exoskeleton assistance for SCI gait training. If successful, RL-Exo could shift exoskeleton rehabilitation from one-size-fits-all to continuously adaptive, patient-centered assistance, maximizing recovery potential while minimizing user effort and fall risk.