Encouraging and Evaluating Embodied Exploration
So Yeon Min 1
1Carnegie Mellon University 2Allen Institute for Artificial Intelligence
Excalibur is a new questions answering task for embodied AI. It is designed both as an evaluation benchmark and as a training signal for simulated robots exploring 3D environements. Different from previous embodied QA tasks, Excalibur requires agents to explore the environment before seeing the questions (Phase I/II), and allows them to reenter the environment to refine their answers (Phase III/IV). It is also more realistic with low-level manipulation actions, and more challenging with procedurally generated scenes and questions. Please first enjoy the following one minute walkthrough video to get a quick overview of the task.