Learning Human Search Behavior from Egocentric Visual Inputs

Human beings can easily search for objects in complex environments. A recent paper proposes a virtual human character that is able to look for any randomly placed objects in a 3D scene using egocentric vision and locomotion capability.

Deep reinforcement learning is used to train the search policy, which determines the movement and gaze direction at every step. A motion synthesis module uses an online replanning method to transfer a model-agnostic policy to a particular robot model and synthesizes the kinematic motion.

Image credit: NASA

During the tests with both human character and wheel-based robot, the character was able to find small objects such as a pair of glasses in a large space even then they were occluded by furniture or placed inside the cabinets. Also, it is shown that enabling head movements improves performance. This study gives hope to further robotic applications.

“Looking for things” is a mundane but critical task we repeatedly carry on in our daily life. We introduce a method to develop a human character capable of searching for a randomly located target object in a detailed 3D scene using its locomotion capability and egocentric vision perception represented as RGBD images. By depriving the privileged 3D information from the human character, it is forced to move and look around simultaneously to account for the restricted sensing capability, resulting in natural navigation and search behaviors. Our method consists of two components: 1) a search control policy based on an abstract character model, and 2) an online replanning control module for synthesizing detailed kinematic motion based on the trajectories planned by the search policy. We demonstrate that the combined techniques enable the character to effectively find often occluded household items in indoor environments. The same search policy can be applied to different full-body characters without the need for retraining. We evaluate our method quantitatively by testing it on randomly generated scenarios. Our work is a first step toward creating intelligent virtual agents with humanlike behaviors driven by onboard sensors, paving the road toward future robotic applications.

Link: https://arxiv.org/abs/2011.03618