Policy Pre-training for Autonomous Driving via Self-supervised Geometric Modeling

TL;DR

The paper tackles sample inefficiency in end-to-end autonomous driving by introducing PPGeo, a fully self-supervised two-stage pretraining framework that learns policy-relevant representations from unlabeled driving videos. It leverages dense geometric modeling to predict ego-motion from frame pairs (Stage One) and then trains a single-frame visual encoder to forecast future ego-motion (Stage Two), with depth and intrinsics fixed. Across CARLA navigation, Leaderboard, RL, and nuScenes planning tasks, PPGeo outperforms a range of baselines, especially with limited labeled data, and depth/odometry training benefits emerge as a useful byproduct on real-world data. The approach yields driving-focused representations concentrated on relevant cues such as lanes and traffic signals, offering improved sample efficiency and generalization for visuomotor driving systems.

Abstract

Witnessing the impressive achievements of pre-training techniques on large-scale data in the field of computer vision and natural language processing, we wonder whether this idea could be adapted in a grab-and-go spirit, and mitigate the sample inefficiency problem for visuomotor driving. Given the highly dynamic and variant nature of the input, the visuomotor driving task inherently lacks view and translation invariance, and the visual input contains massive irrelevant information for decision making, resulting in predominant pre-training approaches from general vision less suitable for the autonomous driving task. To this end, we propose PPGeo (Policy Pre-training via Geometric modeling), an intuitive and straightforward fully self-supervised framework curated for the policy pretraining in visuomotor driving. We aim at learning policy representations as a powerful abstraction by modeling 3D geometric scenes on large-scale unlabeled and uncalibrated YouTube driving videos. The proposed PPGeo is performed in two stages to support effective self-supervised training. In the first stage, the geometric modeling framework generates pose and depth predictions simultaneously, with two consecutive frames as input. In the second stage, the visual encoder learns driving policy representation by predicting the future ego-motion and optimizing with the photometric error based on current visual observation only. As such, the pre-trained visual encoder is equipped with rich driving policy related representations and thereby competent for multiple visuomotor driving tasks. Extensive experiments covering a wide span of challenging scenarios have demonstrated the superiority of our proposed approach, where improvements range from 2% to even over 100% with very limited data.

Figures (7)

• Figure 1: Uniqueness of visuomotor driving policy learning. The planned trajectory is shown as red points. (a) static obstacles and background buildings (objects in yellow rectangles) are irrelevant to the driving decision; (b) the traffic signal in the visual input (marked with the green box) is extremely difficult to recognize and yet deterministic for control outputs; (c) the pre-trained visual encoder has to be robust to different light and weather conditions. Photo credit from caesar2020nuscenes.
• Figure 2: Overview of PPGeo. (a) We focus on pre-training an effective visual encoder to encode driving policy related information by predicting ego-motion based on single frame input (a.2 Stage Two). As achieving such a goal without labels is non-trivial, the visual encoder is obtained with the aid of a preceding procedure (a.1 Stage One) with temporal inputs and two sub-networks (pose and depth). In this illustrative example, the ego-vehicle needs to take action of STOP. The ego-motion in (a.1) is inferred by judging two consecutive frames barely change; whilst the ego-motion in (a.2) is predicted based on single visual input - focusing on driving policy related information. As such, the visual encoder could be fine-tuned and applied to a wide span of downstream tasks in (b).
• Figure 3: Learning curves of the RL agents using PPGeo and three other best pre-training baselines. Left: the pre-trained visual encoder is jointly fine-tuned during RL training; Right: the visual encoder is frozen during RL training. The episode return is the mean with standard deviation in shade across three runs with different random seeds.
• Figure 4: Eigen-Cam eigencam activation maps of the learned representation from different pre-training methods on the driving video data.
• Figure 5: Maps of Town01, Town02, and Town05.