Decentralized Structural-RNN for Robot Crowd Navigation with Deep Reinforcement Learning

Abstract

Persons with visual impairments (PwVI) have difficulties understanding and navigating spaces around them. Current wayfinding technologies either focus solely on navigation or provide limited communication about the environment. Motivated by recent advances in visual-language grounding and semantic navigation, we propose DRAGON, a guiding robot powered by a dialogue system and the ability to associate the environment with natural language. By understanding the commands from the user, DRAGON is able to guide the user to the desired landmarks on the map, describe the environment, and answer questions from visual observations. Through effective utilization of dialogue, the robot can ground the user's free-form descriptions to landmarks in the environment, and give the user semantic information through spoken language. We conduct a user study with blindfolded participants in an everyday indoor environment. Our results demonstrate that DRAGON is able to communicate with the user smoothly, provide a good guiding experience, and connect users with their surrounding environment in an intuitive manner.

Intention Aware Robot Crowd Navigation with Attention-Based Interaction Graph

“…Another line of work, named Deep V-Learning, first uses supervised learning and then RL to learn a value function for path planning based on known state transitions of all agents [8], [11], [12], [20]. To remove assumptions on state transitions, decentralized structural-RNN (DS-RNN) uses model-free RL to train the robot policy from scratch with RL [13]. To model the interactions between the robot and humans, these RL-based methods use long short-term memory (LSTM) encoders, attention mechanisms, and spatio-temporal graphs.…”

“…Attention mechanisms have been widely applied in sequence-based tasks such as trajectory prediction, crowd navigation, and machine translation [8], [13], [23], [27]- [29]. Vaswani et al propose a self-attention mechanism that achieves state-of-the-art performance in machine translation [30].…”

“…Similarly, in trajectory prediction, attention networks value the importance of each pairwise interaction among participants, which improves the predictions of their future trajectories [8], [23], [27]. In crowd navigation of robots and vehicles, previous works have used attention networks to score the importance of each human neighbor to the controlled agent [8], [13], [31]. Compared with these works that ignore the interactions between humans, our attention networks capture the interactions among all observed agents and thus generalize better in dense and highly interactive human crowds.…”

“…2a, the HH and RH functions are denoted by the blue boxes and red boxes respectively and have parameters that need to be learned. Compared with the previous works that ignore HH edges [6], [8], [10], [13], our method considers the pair-wise interactions among all visible agents and thus scales better in dense and highly interactive crowds.…”

“…Robot navigation in dynamic and crowded environments has received much attention for many years [4]- [8]. Previous works have utilized heuristic methods and learning-based methods to determine the robot's optimal action [6], [8]- [13]. These methods define the humans' personal zones as simplistic circles centered at the humans with fixed radii, which represent the non-traversable regions for the robot, as shown in Fig.…”