GRIT: Fast, Interpretable, and Verifiable Goal Recognition with Learned Decision Trees for Autonomous Driving (2103.06113v3)

Abstract: It is important for autonomous vehicles to have the ability to infer the goals of other vehicles (goal recognition), in order to safely interact with other vehicles and predict their future trajectories. This is a difficult problem, especially in urban environments with interactions between many vehicles. Goal recognition methods must be fast to run in real time and make accurate inferences. As autonomous driving is safety-critical, it is important to have methods which are human interpretable and for which safety can be formally verified. Existing goal recognition methods for autonomous vehicles fail to satisfy all four objectives of being fast, accurate, interpretable and verifiable. We propose Goal Recognition with Interpretable Trees (GRIT), a goal recognition system which achieves these objectives. GRIT makes use of decision trees trained on vehicle trajectory data. We evaluate GRIT on two datasets, showing that GRIT achieved fast inference speed and comparable accuracy to two deep learning baselines, a planning-based goal recognition method, and an ablation of GRIT. We show that the learned trees are human interpretable and demonstrate how properties of GRIT can be formally verified using a satisfiability modulo theories (SMT) solver.

• The paper introduces a novel decision tree approach that delivers real-time goal recognition with accuracy competitive to deep learning methods.
• It utilizes interpretable features from vehicle trajectories and scene contexts to provide clear, logical predictions essential for safety-critical autonomous driving.
• The method enables formal verification via SMT solvers, ensuring that its decision-making process meets stringent regulatory and reliability standards.

Overview of GRIT: Goal Recognition with Interpretable Trees for Autonomous Vehicles

The paper presents GRIT, a novel approach for goal recognition in autonomous vehicles, leveraging decision trees to meet the essential criteria of speed, accuracy, interpretability, and verifiability. The automation of goal recognition in autonomous driving is crucial for predicting the future trajectories of surrounding vehicles, particularly in urban environments characterized by dense and multifaceted interactions. While existing methods often fall short in integrating these four objectives cohesively, GRIT addresses this gap using a methodology grounded in decision tree learning.

The proposed system utilizes decision trees trained on vehicle trajectory data to generate real-time, interpretable predictions about the goals of other vehicles. By using decision trees, the authors achieve a balance between computational efficiency and human interpretability, which is often lacking in deep learning approaches. This balance is crucial given the safety-critical nature of autonomous driving, where systems must not only be accurate but also provide explanations for their decisions that are comprehensible to humans.

Key Components of GRIT

GRIT's architecture is designed to infer goal probabilities efficiently. The method involves generating a set of possible goals for each vehicle and extracting a feature vector based on observed trajectories and the static scene information. Decision trees then infer the likelihood of each goal, and these likelihoods are converted into a Bayesian posterior probability distribution over potential goals.

1. Goal Generation: Possible goals for each vehicle are generated considering the local road layout and the vehicle's current state.
2. Feature Extraction: Features such as path length to goal, lane correctness, current speed, and other traffic-related variables are extracted. These are chosen for their interpretability and relevance to autonomous driving scenarios.
3. Decision Trees: The core of the method, decision trees are trained to balance between complexity and interpretability. The structure of the trees facilitates understanding by representing decisions as clear, logical steps, which can be further encoded into propositional logic for verification.
4. Verification: Unlike deep learning methods, the logical and structured nature of decision trees allows for a formal verification process using satisfiability modulo theories (SMT) solvers. This capability is pivotal in a field where the verification of safety-critical systems is paramount.

Evaluation and Outcomes

GRIT is evaluated on datasets from urban driving scenarios and compared to baseline methods, including a deep learning model and a planning-based method. The results demonstrate that GRIT achieves comparable accuracy to deep learning approaches while being significantly more interpretable and verifiable. Specifically, GRIT maintains real-time performance while facilitating a straightforward explanation of its predictions, thus adhering to regulatory requirements such as the "right to explanation."

1. Accuracy and Speed: GRIT achieves a prediction accuracy that is competitive with more complex models while ensuring inference speeds that support real-time applications.
2. Interpretability: The trees learned by GRIT offer a high degree of interpretability, providing insights into decision-making processes, crucial for building trust in autonomous systems.
3. Robust Verification: Through the use of SMT solvers, GRIT's outputs can be formally verified, allowing the establishment of safety guarantees which are difficult to derive from neural networks.

Implications and Future Directions

GRIT sets a new standard in the development of goal recognition systems for autonomous vehicles by emphasizing the need for models that integrate speed, accuracy, interpretability, and verifiability. The decision tree-based approach opens new pathways for developing autonomous systems that are not only reliable but also transparent and accountable.

The implications of this work extend into various facets of AI development for autonomous systems. The modularity of GRIT offers potential adaptability to open-world driving scenarios. Future work might explore integrating knowledge distillation from deep networks to enhance decision trees further, potentially yielding even higher accuracy without sacrificing their beneficial properties. Additionally, the model could be expanded to handle occlusions and more complex urban driving conditions.

In conclusion, GRIT represents a significant step forward in designing goal recognition systems that align technical performance with crucial aspects of safety and interpretability, setting a benchmark for future research in autonomous driving.