GAD-Generative Learning for HD Map-Free Autonomous Driving

TL;DR

The paper addresses the planning bottleneck in autonomous driving by introducing GAD, a data-driven, HD map-free framework that integrates prediction and planning on edge devices. It combines a two-stream BEV rasterization and vectorized instance representation with a generator-evaluator paradigm, including lattice, imitation, and GAN-based trajectory generation guided by a max-margin cost-volume evaluator and a safety layer. Key contributions include an industry-grade, scalable prediction-planning approach validated on factory-ready hardware in urban Shanghai, a multi-modal planning strategy that surpasses imitation-only methods, and rigorous closed-loop evaluation demonstrating robustness beyond offline metrics. The work advances practical deployment by offering interpretable, tunable planning that can adapt to real-world variability while maintaining safety and performance for mass production autonomous systems.

Abstract

Deep-learning-based techniques have been widely adopted for autonomous driving software stacks for mass production in recent years, focusing primarily on perception modules, with some work extending this method to prediction modules. However, the downstream planning and control modules are still designed with hefty handcrafted rules, dominated by optimization-based methods such as quadratic programming or model predictive control. This results in a performance bottleneck for autonomous driving systems in that corner cases simply cannot be solved by enumerating hand-crafted rules. We present a deep-learning-based approach that brings prediction, decision, and planning modules together with the attempt to overcome the rule-based methods' deficiency in real-world applications of autonomous driving, especially for urban scenes. The DNN model we proposed is solely trained with 10 hours of human driver data, and it supports all mass-production ADAS features available on the market to date. This method is deployed onto a Jiyue test car with no modification to its factory-ready sensor set and compute platform. the feasibility, usability, and commercial potential are demonstrated in this article.

Figures (16)

• Figure 1: Comparison on the design of model-based planning. (a)Imitation planner with single trajectory output (hu2023_uniadjiang2023vadrenz2022plantvitelli2022safetynet). (b)Sample a bunch of candidates and evaluate with cost volume(zeng2021endtoendcasas2021mp3sadat2020perceivehu2022st). (c)Proposed generative method, in which the data-driven generator gets feedback from the evaluator.
• Figure 2: The Model Structure of GAD initiates with two streams, rasterization and vectorization, to encode scene and instance-level features; the rasterization branch uses convolution layers to derive the scene embedding, which is decoded into cost volumes and prediction grid maps, while the scene embedding is cropped and combined with the instance encoding from the vectorization branch to form $f_{agent}$ which decoded into ego trajectory and agents' future movements.
• Figure 3: Multi-stage Prediction
• Figure 4: Visualization of comparative results. (a)the predicted trajectory in Ego-ENU closely follows the route. (b)trajectory in the fixed-oriented coordinate system deviates outwards. (c)trajectory in the fixed-oriented coordinate system with feature normalization exhibits proximity to the route.
• Figure 5: Route visualization. (a) lane center line is served as route in cruising scenarios. (b) is the turning scenarios, where historical commuting trajectories are served as route, and occupancy provides road boundary information.