Diffusion-Based Planning for Autonomous Driving with Flexible Guidance
Authors: Yinan Zheng, Ruiming Liang, Kexin ZHENG, Jinliang Zheng, Liyuan Mao, Jianxiong Li, Weihao Gu, Rui Ai, Shengbo Li, Xianyuan Zhan, Jingjing Liu
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Evaluations on the large-scale real-world autonomous planning benchmark nu Plan and our newly collected 200-hour delivery-vehicle driving dataset demonstrate that Diffusion Planner achieves state-of-the-art closed-loop performance with robust transferability in diverse driving styles. 5 EXPERIMENTS |
| Researcher Affiliation | Collaboration | 1 Tsinghua University 2 Institute of Automation, Chinese Academy of Sciences 3 The Chinese University of Hong Kong 4 Shanghai Jiao Tong University 5 HAOMO.AI 6 Shanghai Artificial Intelligence Laboratory |
| Pseudocode | No | No section or figure explicitly labeled "Pseudocode" or "Algorithm" was found in the paper. The methodology is described in narrative text and mathematical formulations. |
| Open Source Code | No | Project website: https://zhengyinan-air.github.io/Diffusion-Planner/. We have collected and evaluated a new 200-hour delivery-vehicle dataset, which is compatible with the nu Plan framework, and we will open-source it. |
| Open Datasets | Yes | Evaluations on the large-scale real-world autonomous planning benchmark, nu Plan (Caesar et al., 2021), to compare Diffusion Planner with other stateof-the-art planning methods. |
| Dataset Splits | Yes | We use the training data from the nu Plan dataset and sample 1 million scenarios for our training set. The Val14 (Dauner et al., 2023b), Test14, and Test14-hard benchmarks (Cheng et al., 2023) are utilized, with all experimental results tested in both closed-loop non-reactive and reactive modes. |
| Hardware Specification | Yes | Training was conducted using 8 NVIDIA A100 80GB GPUs, with a batch size of 2048 over 500 epochs, with a 5-epoch warmup phase. ...the model achieves an inference frequency of 20 Hz on a single A6000 GPU. |
| Software Dependencies | No | The paper mentions using "Adam W optimizer" and "DPM-Solver++" but does not provide specific version numbers for these or any other software dependencies. |
| Experiment Setup | Yes | Training was conducted using 8 NVIDIA A100 80GB GPUs, with a batch size of 2048 over 500 epochs, with a 5-epoch warmup phase. We use Adam W optimizer with a learning rate of 5e 4. We report the detailed setup in Table 5. |