Learning-Deep-Learning

MP3: A Unified Model to Map, Perceive, Predict and Plan

September 2021

tl;dr: Mapless driving with lidar.

Overall impression

This paper continues the line of work of end-to-end self-driving by Raquel’s team. It is heavily inspired by NMP. They both enumerate trajectories based on a learned cost function evaluator of trajectories, rather than generated by a model.

Mapless driving can 1) serve as the fail-safe in the case of localization failures or outdated maps, and 2) potentially unlock self-driving at scale at a much lower cost.

Without map, the search space to plan a safe maneuver from the SDV goes from narrow envelopes around the lane center lines to the full set of dynamically feasible trajectories.

Key ideas

Technical details

Notes