Everyday activities that challenge the capacities of robots.
A primary goal of embodied AI research is to develop intelligent agents that can assist humans in their everyday lives in activities like washing dishes or cleaning floors. While recent years we have seen excellent new benchmarks that helped the field, they are usually restricted to a few tasks, short horizon or in simplified scenarios. Solving real-world challenges requires a new benchmark with realistic, diverse, and complex activities. We present BEHAVIOR, a benchmark with the 100 household activities that represent a new challenge for embodied AI solutions.
BEHAVIOR is a challenge in simulation where embodied agents make continuous full-body control decisions based on sensor information. Agents need to navigate and manipulate the simulated environment with the goal of acomplishing 100 household activities (see here all the activities in pairs of simulated and real-world images, or here for the complete list of activities). BEHAVIOR tests the ability to perceive the environment, plan, and execute complex long-horizon activities that involve multiple objects, rooms, and state transitions, all with the reproducibility, safety and observability offered by a realistic physics simulation.
The results for this first edition of BEHAVIOR will be presented at ICCV21 in October 17th. Together with the announcement of the ranking, we will hold an exciting workshop on embodied AI interactive problems for long horizon activities, with an exciting set of world-renowed speakers from the field….