Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Rethinking Closed-loop Planning Framework for Imitation-based Model Integrating Prediction and Planning

Jiayu Guo, Mingyue Feng, Pengfei Zhu, Chengjun Li, Jian Pu

TL;DR

This paper tackles the challenge of understanding and reducing accumulative errors in closed-loop neural prediction-planning models for autonomous driving. It introduces a two-mode closed-loop framework that jointly employs Motion Prediction and Planning (MPP) and Conditional Motion Prediction (CMP) without altering network architecture, and it implements adaptive scheduling to decouple planning frequency from safety monitoring. Through extensive closed-loop experiments on the nuPlan dataset and Val14 benchmark, the authors demonstrate improved local planning feasibility and safety, with CMP reducing conservative behavior while sustaining efficiency. The work provides a principled approach to integrate CMP into safety monitoring, showing enhanced performance and stability in diverse driving scenarios, and offers practical implications for trustworthy autonomous systems.

Abstract

In recent years, the integration of prediction and planning through neural networks has received substantial attention. Despite extensive studies on it, there is a noticeable gap in understanding the operation of such models within a closed-loop planning setting. To bridge this gap, we propose a novel closed-loop planning framework compatible with neural networks engaged in joint prediction and planning. The framework contains two running modes, namely planning and safety monitoring, wherein the neural network performs Motion Prediction and Planning (MPP) and Conditional Motion Prediction (CMP) correspondingly without altering architecture. We evaluate the efficacy of our framework using the nuPlan dataset and its simulator, conducting closed-loop experiments across diverse scenarios. The results demonstrate that the proposed framework ensures the feasibility and local stability of the planning process while maintaining safety with CMP safety monitoring. Compared to other learning-based methods, our approach achieves substantial improvement.

Rethinking Closed-loop Planning Framework for Imitation-based Model Integrating Prediction and Planning

TL;DR

This paper tackles the challenge of understanding and reducing accumulative errors in closed-loop neural prediction-planning models for autonomous driving. It introduces a two-mode closed-loop framework that jointly employs Motion Prediction and Planning (MPP) and Conditional Motion Prediction (CMP) without altering network architecture, and it implements adaptive scheduling to decouple planning frequency from safety monitoring. Through extensive closed-loop experiments on the nuPlan dataset and Val14 benchmark, the authors demonstrate improved local planning feasibility and safety, with CMP reducing conservative behavior while sustaining efficiency. The work provides a principled approach to integrate CMP into safety monitoring, showing enhanced performance and stability in diverse driving scenarios, and offers practical implications for trustworthy autonomous systems.

Abstract

In recent years, the integration of prediction and planning through neural networks has received substantial attention. Despite extensive studies on it, there is a noticeable gap in understanding the operation of such models within a closed-loop planning setting. To bridge this gap, we propose a novel closed-loop planning framework compatible with neural networks engaged in joint prediction and planning. The framework contains two running modes, namely planning and safety monitoring, wherein the neural network performs Motion Prediction and Planning (MPP) and Conditional Motion Prediction (CMP) correspondingly without altering architecture. We evaluate the efficacy of our framework using the nuPlan dataset and its simulator, conducting closed-loop experiments across diverse scenarios. The results demonstrate that the proposed framework ensures the feasibility and local stability of the planning process while maintaining safety with CMP safety monitoring. Compared to other learning-based methods, our approach achieves substantial improvement.
Paper Structure (31 sections, 16 equations, 5 figures, 3 tables)

This paper contains 31 sections, 16 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Neural Planning with Prediction
  4. Conditional Motion Prediction
  5. Method
  6. Problem Formulation
  7. Achitecture of Neural Network
  8. Scenario Backbone.
  9. Extracting agent-centric interaction.
  10. Temporal aggregation.
  11. Message passing.
  12. Interactive Autogressive Decoder.
  13. Iterative interaction.
  14. Local fusion.
  15. Closed-loop Planning Framework.
  16. ...and 16 more sections

Figures (5)

  • Figure 1: The neural network architecture comprises the scenario backbone and the interactive autoregressive decoder.
  • Figure 2: A closed-loop planning framework for our neural network, featuring a planning mode driven by Motion Prediction and Planning (MPP) and a safety monitoring mode with Conditional Motion Prediction (CMP).
  • Figure 3: Temporal workflow involves adaptive and non-adaptive scheduling. In non-adaptive scheduling, the replanning time interval $T_{plan}$ remains fixed, requiring a high frequency of MPP to ensure safety. Conversely, adaptive scheduling decouples planning driven by MPP and safety monitor with CMP. This allows for a longer $T_{plan}$ while still ensuring safety.
  • Figure 4: Trends of metrics over $T_{plan}$ for NR-CL simulations in Val14 for adaptive and non-adaptive scheduling. The metrics includes the composite score and ratios of collision and off-road
  • Figure 5: Case study. In Case 1, a comparison is made between the utilization of MPP and CMP in safety monitoring when the Ego Vehicle (EV) plans to change lanes. The MPP predicts a collision between the focal SV and the EV, while the CMP predicts the SV to yield. In Case 2, EV replans when changes lane, leads to unnecessary and continuous lane-changing maneuvers. When applying adaptive longer $T_{plan}$, the EV replans to keep lane.