Deep Occupancy-Predictive Representations for Autonomous Driving

Abstract-- Manually specifying features that capture the diversity in traffic environments is impractical. We show that our approach significantly improves the downstream performance of a reinforcement learning agent operating in urban traffic environments. Figure 1: Our spatio-temporal representation model learns a continuous parameterization of the probabilistic occupancy map ô(s, t). To alleviate lack of a canonical ordering of other traffic participants is the lossy nature of compressive graph encoding, we propose incompatible with fixed-sized feature vectors. Second, the a novel occupancy prediction framework that constrains the diversity in road networks in terms of geospatial topology decoding space in accordance with a priori known physical complicates specifying a universal map representation [3].