Humans either over-rely (drive a Tesla while sleeping) or under rely (interfering) with robot’s actions. human and robot interactions may depend on entire history trust is a proxy for the full interaction history the human’s policy must be modeled by th robot trust is demonstrated through real-world experimentation Formulation Add two variable Trust: \theta_{t}, the robot’s ability to succeed in a task Performance: e_{t+1}, success or failure in attempting a task the trust model probabilities for model’s correct modeling of humans are low: high variance between participants. Trust Dynamics models human’s trust in the robot as a linear gaussian. Human Model Results Sadly, the system didn’t actually increase in trust score, but the performance was better through lower human intervention.