Int2Planner: An Intention-based Multi-modal Motion Planner for Integrated Prediction and Planning

TL;DR

Int2Planner tackles uncertainty in autonomous driving by tightly integrating prediction and planning through route-based intention points. A three-part architecture combines a Context Encoder, Route Encoder, and Transformer-based Trajectory Generator to produce multi-modal planning trajectories for multiple intention points, refined over $K$ decoding iterations. Empirical results on a private urban dataset and the nuPlan benchmark show state-of-the-art planning performance, with clear gains from route intention points and integrated prediction, and real-world deployments demonstrate practical viability. The work highlights route-informed intention as a principled mechanism to constrain uncertainty and enable safe, reactive autonomous driving in complex traffic.

Abstract

Motion planning is a critical module in autonomous driving, with the primary challenge of uncertainty caused by interactions with other participants. As most previous methods treat prediction and planning as separate tasks, it is difficult to model these interactions. Furthermore, since the route path navigates ego vehicles to a predefined destination, it provides relatively stable intentions for ego vehicles and helps constrain uncertainty. On this basis, we construct Int2Planner, an \textbf{Int}ention-based \textbf{Int}egrated motion \textbf{Planner} achieves multi-modal planning and prediction. Instead of static intention points, Int2Planner utilizes route intention points for ego vehicles and generates corresponding planning trajectories for each intention point to facilitate multi-modal planning. The experiments on the private dataset and the public nuPlan benchmark show the effectiveness of route intention points, and Int2Planner achieves state-of-the-art performance. We also deploy it in real-world vehicles and have conducted autonomous driving for hundreds of kilometers in urban areas. It further verifies that Int2Planner can continuously interact with the traffic environment. Code will be avaliable at https://github.com/cxlz/Int2Planner.

Figures (8)

• Figure 1: The overall framework of Int2Planner. (a) denotes the Context Encoder module, which encodes agent states and HD map information into context embedding, (b) denotes the Route Encoder module, which encodes route information into route embedding and (c) denotes the Trajectory Generator module, which optimizes future trajectories with $K$ iterations, (d) denotes the detailed structure of the Trajectory Decoder layer and "$\oplus$" denotes a concatenation operation of tensors.
• Figure 2: Explanation of route intention points sampling.
• Figure 3: Qualitative results. (a)-(d) Primary and secondary intention points are marked with red and blue "$\star$". Intention points with the highest confidence are marked larger, along with their confidence scores. (e)-(g) The three most confident planning trajectories are plotted with intention points and confidence scores.
• Figure 4: The distribution of planning confidence score.
• Figure 5: Real-world tests in urban areas. The front view image are combined with four surrounding view images and the output planning trajectories are projected as yellow lines.