A survey on deep reinforcement learning for human-robot interaction

Abstract

Deep reinforcement learning (DRL) is reshaping human-robot interaction (HRI) control beyond fixed admittance/impedance laws. This survey bridges classical HRI control and DRL, then organises recent work into four paradigms: physical HRI via admittance shaping, learning from demonstration and feedback, multi-agent and shared-autonomy control, and safe, stability-certified reinforcement learning (RL). We further discuss emerging directions, language-conditioned policies, digital-twin-based sim-to-real transfer, and residual RL, and argue that, while conformal-prediction bounds are now routinely combined with set-invariance guarantees such as barrier functions in HRI, their pairing with Lyapunov/uniform ultimate boundedness (UUB) stability guarantees remains obstructed by the loss of exchangeability in closed loop, which we identify as the field’s least developed dimension. We close with open challenges in sample efficiency, non-stationary human behaviour, generalisation across users, and benchmarking.