Planning by Simulation: Motion Planning with Learning-based Parallel Scenario Prediction for Autonomous Driving

TL;DR

A novel motion Planning approach by Simulation with learning-based parallel scenario prediction (PS) that effectively achieves parallel ego vehicle planning and adopts MCTS as the foundation to explore possible future interactions encoded within the prediction network.

Abstract

Planning safe trajectories for autonomous vehicles is essential for operational safety but remains extremely challenging due to the complex interactions among traffic participants. Recent autonomous driving frameworks have focused on improving prediction accuracy to explicitly model these interactions. However, some methods overlook the significant influence of the ego vehicle's planning on the possible trajectories of other agents, which can alter prediction accuracy and lead to unsafe planning decisions. In this paper, we propose a novel motion Planning approach by Simulation with learning-based parallel scenario prediction (PS). PS deduces predictions iteratively based on Monte Carlo Tree Search (MCTS), jointly inferring scenarios that cooperate with the ego vehicle's planning set. Our method simulates possible scenes and calculates their costs after the ego vehicle executes potential actions. To balance and prune unreasonable actions and scenarios, we adopt MCTS as the foundation to explore possible future interactions encoded within the prediction network. Moreover, the query-centric trajectory prediction streamlines our scene generation, enabling a sophisticated framework that captures the mutual influence between other agents' predictions and the ego vehicle's planning. We evaluate our framework on the Argoverse 2 dataset, and the results demonstrate that our approach effectively achieves parallel ego vehicle planning.

Figures (12)

• Figure 1: Illustration of the planning by simulation. First, multiple future scenarios are predicted by our learning-based parallel prediction module, which simulates the ego vehicle and our vehicle simultaneously. The scenario prediction is conducted by a MCTS search. Safe and smooth ego planning are chosen out of multiple possible futures. Compared to pure e2e planning, our framework is more robust thanks to explicit future reasoning. More examples can be found in the video https://youtu.be/_Y9HGzgKfsc.
• Figure 2: Illustration of the proposed trajectory planning framework. Our model takes high-definition maps and agents' historical states as inputs to extract lane polygons and agents' trajectories. Based on these inputs, the model generates ego planning set. The embedded prediction model separates map encoding from agent encoding and outputs high-accuracy future trajectories for the agents. We execute the prediction process in parallel, expanding the search tree layer by layer. After that, the Monte Carlo policy helps prune branches with poor states, and the remaining states become new nodes, quickly encoded into the prediction model. This process repeats until the specified depth is reached, achieving a cyclic interaction between prediction and planning.
• Figure 3: Illustration of planning future scenarios. This reasoning process, which iterates dozens of times, is based on MCTS and is composed of four steps: selection, expansion, simulation, and backpropagation.
• Figure 5: Illustration of how to update other agents' states. Our model encodes history nodes on the branch as $\left [ N,\tau -1,D \right ]$ dimension agents embedding and $\left [M ,D \right ]$ dimension map embedding, where $D$ represents the dimension of encoding hidden layer. When a new node $s_{n=1,...,N} ^{T+\tau }$ expands, temporal attention encodes $s_{n=1,...,N} ^{T+\tau }$ as query and historical agents encoding as key and value. After decoding, the model outputs $k$ future trajectories and their probabilities.
• Figure : (a)