Vision-and-dialog navigation is a recent benchmark for evaluating the AI capabilities of perception, interaction, and decision making. While existing methods developed for this benchmark have demonstrated great successes, they mostly rely on large datasets, where data collection can be a challenge, and the learned policies are not adaptive to domain changes. In this paper, we focus on a new problem, referred to as goal-oriented vision-and-dialog navigation (GVDN), where an agent uses reinforcement learning techniques to compute dialog-navigation policies from trial and error. A robot conducts visual navigation to locate target objects, and can talk to a remote human operator as needed. Our remote human is able to provide guidance on navigation only if the robot correctly conveys its location through dialog. Experiments have been conducted using photo-realistic simulation environments. Results suggest that, our agent outperforms competitive baselines in success rate.

Reinforcement learning (RL) methods have been widely used for learning dialog policies. Sample efficiency, i.e., the efficiency of learning from limited dialog experience, is particularly important in RL-based dialog policy learning, because interacting with people is costly and low-quality dialog policies produce very poor user experience. In this paper, we develop LHUA (Learning with Hindsight, User modeling, and Adaptation) that, for the first time, enables dialog agents to adaptively learn with hindsight from both simulated and real users. Simulation and hindsight provide the dialog agent with more experience and more (positive) reinforcement respectively. Experimental results suggest that LHUA outperforms competitive baselines from the literature, including its no-simulation, no-adaptation, and no-hindsight counterparts.