Scale human‑robot interaction training to more difficult tasks

Ascertain whether and how direct real‑world reinforcement learning and high‑fidelity human behavior simulation can be scaled to complex, safety‑critical human‑robot interaction tasks, and develop the necessary methods and evaluation procedures to do so.

Background

Two main approaches for HRI training are discussed: direct learning alongside humans in the real world, and simulation with learned or scripted human models. These have shown promise on simpler tasks but face challenges of safety, sample efficiency, and realism.

The authors explicitly note uncertainty about the scalability of these approaches to more complex HRI scenarios, underscoring the need for methods and infrastructure that can handle increased task difficulty and safety requirements.