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Mapless driving can 1) serve as the fail-safe in the case of localization failures or outdated maps, and 2) potentially unlock self-driving at scale at a much lower cost.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
- Online map
- drivable area
- reachable lanes (unsigned distance to the closest reachable lane centerline)
- intersections (traffic is controlled via traffic lights or traffic signs)
- dynamic occupancy field = initial occupancy + temporal motion field (11 x 0.5 = 5 s)
- which space is occupied by dynamic objects and how do they move over time.
- computation is agnostic to the number of objects.
- usually motion forecasting involve unsafe discrete decisions such as confidence thresholding and NMS.
- P3 proposed the idea of dynamic occupancy field
- Motion planning has 3 goals: safe, comfortable and progressing toward the goal
- Trajectory sampling
- sampled from large scale dataset of real trajectories
- To generate continuous velocity and steering, instead of directly using the trajectories, use the acc and steering rate profiles to rollout a bicycle model from the initial SDV state.
- Route prediction
- The network takes in a driving command and output a dense probability map R in BEV.
- A driving command is characterized by a tuple (a, d) where a = {keep, left, right} and d is longitudinal distance to action. This is similar to command given by the GPS
- Trajectory scoring
- The trajectory needs to overlap high with route map R.
- The trajectory needs to stay on the road and avoid encroaching onto the sidewalks or curbs.
- The trajectory needs to be safe, and gets penalized for overlapping with dynamic grids.
- The trajectory needs to be comfortable and gets penalized for jerk, lateral accel, curvature and curvature change rates.
Technical details
- Lidar input formulation follows that of FaF, intentNet and Pixor, with 0.2 m/voxel, with motion compensation.
- Two-stage training: train online mapping, dynamic occupancy field and routing.
Notes