Active Probing with Multimodal Predictions for Motion Planning

Darshan Gadginmath, Farhad Nawaz, Minjun Sung, Faizan M. Tariq, Sangjae Bae, David Isele, Fabio Pasqualetti, Jovin D'sa

TL;DR

This paper introduces a novel framework for autonomous vehicle motion planning that tackles the challenge of uncertain predictions from other vehicles. It combines multimodal trajectory predictions with an "active probing" mechanism, allowing the ego vehicle to strategically interact with others to better infer their intentions and ensure safer, more efficient navigation.

What Problem Does This Paper Address?

Autonomous vehicles operate in dynamic environments filled with other road users whose future actions are inherently uncertain. While current prediction models can generate multiple possible trajectories (multimodal predictions) for these agents, they often struggle with out-of-distribution scenarios and variations in driving styles due to limitations in training data. This uncertainty can lead to overly conservative planning, causing traffic congestion, or worse, unsafe situations. Existing approaches often rely on passive observation, waiting for other agents to reveal their intentions, which is inefficient. This paper addresses the need for a proactive approach that actively reduces prediction ambiguity and handles multimodal uncertainties to enable robust and efficient motion planning.

Methods

The core of the proposed framework is a unified approach that integrates three key components:

  • Multimodal Predictions with Uncertainty: The system leverages existing multimodal prediction models but introduces a novel risk metric that rigorously quantifies uncertainty. This metric is proven to be analytically tractable and always finite when prediction uncertainties follow a Gaussian mixture distribution.
  • Active Probing for Behavior Inference: This is the paper's main innovation. Instead of passively observing, the ego vehicle takes "proactive strategic actions" (e.g., slight acceleration or steering adjustments) to elicit responses from other agents. These actions are designed to reveal the other agents' intentions and behavioral parameters (e.g., aggressiveness, defensiveness) without being overly conservative or disruptive. The framework uses a modified Boltzmann model to infer these parameters based on the predicted likelihoods and the ego vehicle's actions.
  • Unified Optimization Framework: The system optimizes a combined objective function that balances three competing goals: maximizing utility (following a reference trajectory, comfort), ensuring safety (minimizing risk), and maximizing information gain about other agents' behaviors through active probing. This is achieved within a Model Predictive Control (MPC) framework, where the plan is re-evaluated at each time step.

Datasets

The paper evaluates its framework using the MetaDrive simulation environment. MetaDrive is highlighted for its realistic vehicle dynamics and diverse traffic scenarios, making it suitable for testing autonomous navigation in complex situations.

Key Insights and Ideas

  • Proactive Information Gathering is Key: The most significant insight is that actively probing other agents, rather than passively waiting, can dramatically improve the ego vehicle's understanding of their intentions and lead to more efficient and safer planning.
  • Bridging the Gap Between Theory and Practice: The framework's computational tractability, especially with the closed-form risk solution and efficient scaling with the number of agents and prediction modes, is crucial for real-world deployment, addressing a major limitation of prior work.
  • Risk as a Guide for Probing: The paper elegantly links risk assessment with active probing. When the risk of a particular prediction is high, probing is limited to avoid unsafe actions, ensuring that information gathering doesn't compromise safety.
  • Behavioral Parameters as a Unified Representation: By modeling agent behavior through a set of parameters (e.g., weights for velocity matching, safe distance, lane positioning), the framework can adapt to a wide spectrum of driving styles, from aggressive to defensive.

Key Results

The proposed active probing approach demonstrated superior performance in both lane change and intersection scenarios within the MetaDrive environment.

  • Lane Change Scenario: The algorithm achieved a 98% success rate, significantly outperforming a state-of-the-art chance-constrained planner (CC-MPC) at 62% and a no-probing variant at 82%. It also resulted in faster merging times.
  • Intersection Scenario: The probing algorithm again led in success rate (96%), with a 0% collision rate, compared to CC-MPC (89% success, 3% collision) and no-probing (93% success, 0.5% collision).
  • Qualitatively, the algorithm successfully identified and exploited defensive behaviors to create merging opportunities and executed conservative merges when faced with aggressive traffic.

Limitations and Future Work

  • Behavioral Model Simplification: The paper acknowledges that the current behavioral model for other agents is a simplification. Future work could explore more general and complex behavior models.
  • Computational Optimizations: While the framework is computationally efficient, further optimizations are suggested, particularly in prediction filtering and simplified speed planning.
  • Real-World Deployment: The current evaluation is simulation-based. Future work would involve validating the approach in real-world driving scenarios.

Summary updated

Abstract

Navigation in dynamic environments requires autonomous systems to reason about uncertainties in the behavior of other agents. In this paper, we introduce a unified framework that combines trajectory planning with multimodal predictions and active probing to enhance decision-making under uncertainty. We develop a novel risk metric that seamlessly integrates multimodal prediction uncertainties through mixture models. When these uncertainties follow a Gaussian mixture distribution, we prove that our risk metric admits a closed-form solution, and is always finite, thus ensuring analytical tractability. To reduce prediction ambiguity, we incorporate an active probing mechanism that strategically selects actions to improve its estimates of behavioral parameters of other agents, while simultaneously handling multimodal uncertainties. We extensively evaluate our framework in autonomous navigation scenarios using the MetaDrive simulation environment. Results demonstrate that our active probing approach successfully navigates complex traffic scenarios with uncertain predictions. Additionally, our framework shows robust performance across diverse traffic agent behavior models, indicating its broad applicability to real-world autonomous navigation challenges. Code and videos are available at https://darshangm.github.io/papers/active-probing-multimodal-predictions/.
BibTeX
@inproceedings{gadginmath2025active,
  author = {Darshan Gadginmath and Farhad Nawaz and Minjun Sung and Faizan M. Tariq and Sangjae Bae and David Isele and Fabio Pasqualetti and Jovin D'sa},
  title = {{Active Probing with Multimodal Predictions for Motion Planning}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2025},
  doi = {10.1109/iros60139.2025.11246891},
}
Figure