As the cost of rehabilitation robots decreases, it is increasingly common for rehabilitation clinics to own multiple robots. Patients can use such robots to exercise together in a group setting, tentatively dubbed a robotic rehabilitation gym. Such gyms have potential as a new rehabilitation delivery model, as they could increase patient motivation through social interaction and reduce the number of therapists needed per patient. However, such gyms may benefit from additional software support to avoid overloading the supervising therapist and allow them to focus on the most critical aspects of therapy. This paper thus presents our ongoing work on intelligent dynamic patient-robot assignment in a simulated robotic gym via either stochastic optimization or learning from a human expert. Furthermore, we discuss possible next steps for actual real-world studies of robotic gyms.