arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2412.08643cs.CVcs.AIcs.LGcs.RO
收藏

GPD-1: Generative Pre-training for Driving(驾驶生成式预训练)

GPD-1: Generative Pre-training for Driving

Peking University(北京大学);Tsinghua University(清华大学);Phigent Robotics(智芯机器人(音译))

Zixun Xie, Sicheng Zuo, Wenzhao Zheng, Yunpeng Zhang, Dalong Du, Jie Zhou, Jiwen Lu, Shanghang Zhang

AI总结:

该研究提出统一的驾驶生成式预训练模型GPD-1,将自动驾驶建模为token生成问题,通过分层位置tokenizer、地图矢量量化自编码器等设计,在nuPlan数据集预训练后无需微调即可泛化到场景生成、运动规划等多类驾驶任务。

摘要:

对驾驶场景的演化进行建模对自动驾驶系统的评估与决策至关重要。现有大多数方法仅聚焦场景演化的单一维度,例如地图生成、运动预测和轨迹规划。本文提出了一个统一的Generative Pre-training for Driving(GPD-1,驾驶生成式预训练)模型,无需额外微调即可同时完成所有这些任务。我们用自车(ego)、智能体(agent)和地图(map)三类token表示每个场景,并将自动驾驶建模为一个统一的token生成问题。模型采用自回归transformer架构,使用场景级注意力掩码实现场景内的双向交互。针对自车和智能体token,我们提出分层位置tokenizer,可有效编码2D位置和航向角。针对地图token,我们训练了一个地图矢量量化自编码器,可将以自车为中心的语义地图高效压缩为离散token。我们在大规模nuPlan数据集上对GPD-1进行预训练,并开展大量实验验证其有效性。通过不同提示,GPD-1无需微调即可成功泛化到多种任务,包括场景生成、交通仿真、闭环仿真、地图预测和运动规划。代码地址:https://github.com/wzzheng/GPD。

英文摘要:

Modeling the evolutions of driving scenarios is important for the evaluation and decision-making of autonomous driving systems. Most existing methods focus on one aspect of scene evolution such as map generation, motion prediction, and trajectory planning. In this paper, we propose a unified Generative Pre-training for Driving (GPD-1) model to accomplish all these tasks altogether without additional fine-tuning. We represent each scene with ego, agent, and map tokens and formulate autonomous driving as a unified token generation problem. We adopt the autoregressive transformer architecture and use a scene-level attention mask to enable intra-scene bi-directional interactions. For the ego and agent tokens, we propose a hierarchical positional tokenizer to effectively encode both 2D positions and headings. For the map tokens, we train a map vector-quantized autoencoder to efficiently compress ego-centric semantic maps into discrete tokens. We pre-train our GPD-1 on the large-scale nuPlan dataset and conduct extensive experiments to evaluate its effectiveness. With different prompts, our GPD-1 successfully generalizes to various tasks without finetuning, including scene generation, traffic simulation, closed-loop simulation, map prediction, and motion planning. Code: https://github.com/wzzheng/GPD.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑