With the aging of the global population and the emerging wars, the demand for caregiving services continues to increase. A potential solution to this is automation. Specifically, using robots to replace human caregivers in certain daily tasks for elderly, medical and rehabilitation patients, people with disabilities and injuries. In this paper, we consider the problem of learning neural network-based policies for multistage assistive robotics tasks. Neural network policies can significantly simplify the control stack, making the resulting assistive appliances more affordable due to lower requirements to hardware, such as sensors and actuators. However, training an end-to-end neural network policy for a specific assistive task poses several problems. First, the debugging process for such a policy may be challenging due to the multiphased nature of the trajectories. Second, training a few simpler policies for each task stage may be easier and would require less data. Lastly, it is possible to retrain or fine-tune specific stages, or even replace them entirely with a simpler, more traditional planning algorithm. We propose a training method in which we train multiple policies for separate task stages and a stage completion classifier. The policies are trained using a reinforcement learning with prior data to make the approach more flexible and adjustable to engineer’s needs. We evaluate the method on several assistive robotics benchmarks in simulation, demonstrating the better performance against common reinforcement learning and imitation learning approaches.