This repository provides the official implementation for "Diffusion-Based Planning for Autonomous Driving with Flexible Guidance," an ICLR 2025 Oral presentation. It addresses autonomous driving motion planning by leveraging diffusion models, aiming for high performance without relying on extensive refinement steps. The target audience includes researchers and engineers in autonomous driving and robotics.

How It Works

The Diffusion Planner utilizes a DiT-based architecture to fuse noised future vehicle trajectories with conditional information. It jointly models the states of key participants, unifying motion prediction and closed-loop planning into a single future trajectory generation task. This approach allows for fast inference during diffusion sampling, achieving approximately 20Hz for real-time performance.

Quick Start & Requirements

• Installation: Requires conda for environment setup and nuplan-devkit. Clone both repositories and install them using pip install -e . with their respective requirements.txt files.
• Prerequisites: Python 3.9, PyTorch, nuplan-devkit.
• Data: Requires the nuPlan dataset.
• Model: Download checkpoints from Huggingface.
• Links: Arxiv, Project Page, nuplan-devkit

Highlighted Details

• Achieves state-of-the-art closed-loop performance on the nuPlan benchmark, outperforming methods like PLUTO and PlanTF.
• Demonstrates strong results even without refinement steps, a key differentiator.
• Offers flexible guidance capabilities for planning.
• Includes visualizations of future trajectory generation in challenging scenarios.

Maintenance & Community

The project is associated with Tsinghua University and has an upcoming release of additional features, including end-to-end and real-world vehicle integration, and a delivery vehicle dataset.

Licensing & Compatibility

The repository does not explicitly state a license. The project acknowledges inspiration from several open-source projects, including nuplan-devkit, GameFormer-Planner, tuplan_garage, planTF, pluto, StateTransformer, DiT, and dpm-solver.

Limitations & Caveats

The code is still under cleaning and will be released gradually. Some features, such as end-to-end and real-world vehicle integration, are marked as "To Do." The delivery vehicle dataset is pending government approval.