TL;DR
This paper introduces GOFI, an algorithm that jointly infers a vehicle's goals and the presence of occluded factors (like unseen vehicles) to improve long-horizon prediction for autonomous driving. By considering both, GOFI significantly enhances prediction accuracy and leads to safer planning in complex scenarios where traditional methods fail.
What Problem Does This Paper Address?
Autonomous vehicles need to predict the future behavior of other road users to navigate safely. A key challenge is inferring the intentions or "goals" of these other vehicles. While inverse-planning methods can infer goals by assuming rational behavior, they struggle when crucial information is hidden due to occlusions (e.g., a building blocking the view of an oncoming car). In such cases, an autonomous vehicle might incorrectly infer a goal based on incomplete observations, leading to unsafe decisions. For instance, a car slowing down might be interpreted as yielding to the ego-vehicle, when in reality, it's waiting for an occluded vehicle to pass. This paper tackles this problem by developing a method that can simultaneously infer both the goal of an observed vehicle and the presence of any occluded factors influencing its behavior.
Methods
The core of this paper is the Goal and Occluded Factor Inference (GOFI) algorithm. GOFI builds upon existing inverse-planning techniques for goal recognition. The novelty lies in its ability to jointly infer a probabilistic belief over both the possible goals of an observed vehicle and the possible instantiations of occluded factors.
Here's a breakdown of the approach:
- Inverse Planning for Joint Inference: GOFI models the observed vehicle's trajectory as a result of its goal and the presence of occluded factors. It computes the likelihood of the observed trajectory given a specific goal and occluded factor configuration. This likelihood is then used with Bayes' rule to derive a posterior probability distribution over all possible goal-occluded factor combinations.
- Rationality Assumption: Like prior inverse-planning methods, GOFI assumes observed vehicles behave near-rationally to achieve their goals, considering any occluded factors.
- Monte Carlo Tree Search (MCTS) Integration: The inferred probabilistic beliefs from GOFI are then integrated into an MCTS planning framework for the ego-vehicle. During MCTS, the algorithm samples possible goals and occluded factor configurations based on GOFI's output. These sampled states are treated as ground truth for planning, allowing the ego-vehicle to explore actions that are safe under various plausible future scenarios.
- Macro-Actions: For efficient long-horizon planning, MCTS utilizes macro-actions (sequences of maneuvers like "follow lane to exit").
Datasets
Not discussed in the paper. The evaluation uses simulated driving scenarios designed to specifically test the GOFI algorithm's capabilities in the presence of occlusions.
Key Insights and Ideas
- Occlusions Confuse Goal Recognition: The paper clearly demonstrates that ignoring occluded factors can lead to incorrect goal inferences, even if the observed vehicle is acting rationally. This is a critical insight for real-world autonomous driving.
- Joint Inference is Crucial: The core idea is that goals and occluded factors are not independent when explaining observed behavior. They are intertwined, and inferring them together provides a more accurate picture of the world state.
- Observed Behavior as a "Sensor": The behavior of other vehicles can act as an indirect sensor for detecting occluded factors, especially when direct perception is limited.
- Interpretability Matters: By using inverse-planning, GOFI provides interpretable predictions, which is vital for building trust in autonomous systems. The inferred goals and occluded factors can be understood by humans.
Key Results
- Improved Inference Accuracy: GOFI consistently achieved higher posterior probabilities for the true goal and the presence of occluded factors compared to baselines that only considered goals (GR-Only) or only occluded factors (OF-Oracle, which had privileged information).
- Safer Planning: In simulated scenarios requiring occluded factor inference, GOFI-integrated MCTS achieved significantly higher completion rates (fewer collisions) than baselines, including GR-Only. This highlights the practical benefit of accurate inference for safe navigation.
- Complementary to Perception: Even when perception of occluded factors becomes available, GOFI's inference based on behavior can help plan for potential occlusions earlier, leading to safer outcomes.
Limitations and Future Work
- Computational Demand: The paper acknowledges that computing plans for all possible goal-occluded factor combinations can be computationally intensive, especially in scenarios with many vehicles and potential occluded locations. However, they suggest parallel implementation can mitigate this.
- Defining Occluded Factor Space: A limitation is the need to define the set of possible occluded factors and their locations beforehand. Future work aims to address the automatic construction of this set, including identifying occluded locations and their associated behaviors.
- Information Gathering Actions: The authors plan to explore how autonomous vehicles can actively select actions to gather information and resolve uncertainty about occluded factors.
Summary updated