RehabBot: Reinforcement learning environment for upper and lower limb robot-assisted rehabilitation

We present RehabBot, an open-source simulation environment for reinforcement learning (RL) research in robot-assisted upper- and lower-limb rehabilitation. Built on the MuJoCo physics engine, it integrates a torque-controlled UR5e collaborative manipulator with an articulated humanoid patient model to implement physiotherapy-inspired exercises for the shoulder, elbow, hip, and knee. The platform provides a modular RL-ready API with defined continuous observation and action spaces, task-aligned reward functions with safety constraints, and configurable evaluation scenarios. Exercises are formalized as Markov Decision Processes, ensuring compatibility with standard RL libraries. Validation through representative deep RL algorithms demonstrates stable training behaviour and compatibility with standard pipelines. RehabBot offers an extensible framework for reproducible research at the intersection of robotics, rehabilitation, and learning-based control.