Physical Therapy Robot

After mapping out what a robot should say in healthcare conversations, we turned to a harder question: when should a robot act? In post-stroke arm rehabilitation, a robot assistant works alongside a human physiotherapist and a patient. Getting intervention timing wrong doesn't just create awkwardness — it can undermine therapeutic trust or endanger the patient.

We built a Q-learning reinforcement simulation with Hidden Markov Model (HMM) intent recognition to model five collaborative contexts: patient independence, patient needs support, doctor leading, emergency, and collaborative moment. The robot's reward structure was designed so that interrupting a doctor-patient interaction was penalised, while responding to a clear pain signal was given the highest reward of any action. The robot literally learns that sometimes the best thing it can do is nothing.

Privacy-by-design was a core constraint: the system learns from behavioural observation only, no private health data stored. The result is a transparent, auditable AI decision process where healthcare professionals can verify the robot's reasoning at every step.