[Paper Review] Integrating Planning and Interpretable Goal Recognition for Autonomous Driving.

Stefano V. Albrecht, Cillian Brewitt+4 more|arXiv (Cornell University)|Feb 6, 2020

Autonomous Vehicle Technology and Safety26 references4 citations

TL;DR

This paper proposes an integrated planning and interpretable goal recognition system for autonomous driving that uses rational inverse planning to infer other vehicles' goals and Monte Carlo Tree Search (MCTS) to generate near-optimal, explainable maneuvers for the ego vehicle. By sharing a common set of defined maneuvers, the system ensures plan rationality and provides intuitive explanations for predictions and decisions in urban driving scenarios.

ABSTRACT

The ability to predict the intentions and driving trajectories of other vehicles is a key problem for autonomous driving. We propose an integrated planning and prediction system which leverages the computational benefit of using a finite space of maneuvers, and extend the approach to planning and prediction of sequences (plans) of maneuvers via rational inverse planning to recognise the goals of other vehicles. Goal recognition informs a Monte Carlo Tree Search (MCTS) algorithm to plan optimal maneuvers for the ego vehicle. Inverse planning and MCTS utilise a shared set of defined maneuvers to construct plans which are explainable by means of rationality, i.e. plans are optimal in given metrics. Evaluation in simulations of four urban driving scenarios demonstrate the system's ability to robustly recognise the goals of other vehicles while generating near-optimal plans. In each scenario we extract intuitive explanations for the recognised goals and maneuver predictions which justify the system's decisions.

Motivation & Objective

• To improve intention and trajectory prediction of other vehicles in autonomous driving by combining planning and goal recognition.
• To address the challenge of generating explainable, rational maneuvers for the ego vehicle based on inferred goals of surrounding agents.
• To enable robust goal recognition in complex urban driving scenarios using a shared maneuver space across planning and prediction.
• To provide intuitive, rational explanations for both maneuver predictions and ego vehicle decisions through inverse planning.
• To evaluate the system's performance in generating near-optimal plans while maintaining interpretability in simulation.

Proposed method

• Uses a finite set of predefined maneuvers as the basis for both planning and prediction, enabling consistency and interpretability.
• Employs rational inverse planning to infer the goals of other vehicles by identifying the most likely goal that would make their observed maneuvers optimal.
• Integrates goal recognition results into a Monte Carlo Tree Search (MCTS) framework to plan optimal maneuvers for the ego vehicle.
• Ensures all plans are rational by optimizing under defined metrics, making the system's decisions explainable through optimality.
• Constructs plans as sequences of maneuvers, allowing for prediction of complex behavioral sequences rather than single actions.
• Uses shared representations between inverse planning and MCTS to maintain coherence and reduce computational overhead.

Experimental results

Research questions

• RQ1 Can rational inverse planning effectively recognize the goals of other vehicles in diverse urban driving scenarios?
• RQ2 How well can the system predict sequences of maneuvers by other vehicles using goal recognition?
• RQ3 To what extent does integrating goal recognition with MCTS improve the optimality and explainability of ego vehicle planning?
• RQ4 Can the system generate intuitive, rational explanations for its predictions and decisions in complex traffic environments?
• RQ5 How robust is the system in maintaining performance across multiple simulated urban driving scenarios?

Key findings

• The system successfully recognizes the goals of other vehicles in four distinct urban driving scenarios with high robustness.
• Predicted maneuver sequences for other vehicles are consistent with observed behaviors and grounded in rationality.
• The ego vehicle generates near-optimal plans by leveraging recognized goals through MCTS, as validated in simulation.
• The system produces intuitive, human-readable explanations for both goal recognition and maneuver planning decisions.
• The integration of inverse planning and MCTS using a shared maneuver space improves interpretability without sacrificing planning quality.
• The approach demonstrates strong performance across diverse urban scenarios, indicating generalization potential.