Towards Intelligible Human-Robot Interaction: An Active Inference Approach to Occluded Pedestrian Scenarios

TL;DR

A novel framework grounded in Active Inference endows the agent with a human-like, belief-driven mechanism that significantly reduces the collision rate compared to reactive, rule-based, and reinforcement learning baselines, while also exhibiting explainable and human-like driving behavior that reflects the agent's internal belief state.

Abstract

The sudden appearance of occluded pedestrians presents a critical safety challenge in autonomous driving. Conventional rule-based or purely data-driven approaches struggle with the inherent high uncertainty of these long-tail scenarios. To tackle this challenge, we propose a novel framework grounded in Active Inference, which endows the agent with a human-like, belief-driven mechanism. Our framework leverages a Rao-Blackwellized Particle Filter (RBPF) to efficiently estimate the pedestrian's hybrid state. To emulate human-like cognitive processes under uncertainty, we introduce a Conditional Belief Reset mechanism and a Hypothesis Injection technique to explicitly model beliefs about the pedestrian's multiple latent intentions. Planning is achieved via a Cross-Entropy Method (CEM) enhanced Model Predictive Path Integral (MPPI) controller, which synergizes the efficient, iterative search of CEM with the inherent robustness of MPPI. Simulation experiments demonstrate that our approach significantly reduces the collision rate compared to reactive, rule-based, and reinforcement learning (RL) baselines, while also exhibiting explainable and human-like driving behavior that reflects the agent's internal belief state.

Figures (6)

• Figure 1: Reactive vs. Proactive Driving in a Critical Occluded Pedestrian Scenario. (a) A reactive agent ignores the uncertainty from occlusion, fails to slow down, and collides. (b) Our proactive agent reasons about the latent hazard, preemptively reducing speed to ensure a safe interaction with interpretable behaviors.
• Figure 2: Overview of the Active Inference Interaction Framework. Our framework enhances the core perception-action loop (top center) with two key mechanisms to facilitate a explainable interaction. For robust Belief Updating (top-right), Conditional Reset (bottom-left) prevents belief decay from occlusion. For proactive Planning (top-left), Hypothesis Injection (bottom-right) challenges the agent with worst-case scenarios, yielding cautious and belief-driven behavior.
• Figure 3: Impact of Conditional Reset in an Occlusion Scenario. (a) Belief in the pedestrian's existence, $B_t(z_p)=\sum_n w_t^{(n)} \hat{z}_p^{(n)}$, and (b) the agent's speed over time. The agent without Conditional Reset (blue) allows its belief to decay to zero, leading to risky acceleration and a collision. In contrast, the agent with Conditional Reset (orange) maintains its belief, decelerates cautiously, and safely navigates past the revealed pedestrian to reach its goal.
• Figure 4: Effect of Hypothesis Injection Ratio ($\rho_H$) on Behavior. (a) Agent's trajectory and (b) speed profile for different $\rho_H$. A low ratio ($\rho_H=0.0$) leads to insufficient deceleration and a collision. As $\rho_H$ increases, the agent plans a more cautious trajectory with a wider berth (a) by decelerating more significantly (b), thus successfully avoiding the pedestrian.
• Figure 5: Impact of Initial Presence Belief ($B_0(z_p)$) on Behavior. (a) Agent's trajectory and (b) speed profile for different initial beliefs. A zero belief ($B_0(z_p)=0.0$) leads to no evasive action and a collision. As the belief increases, the agent decelerates earlier and more significantly, ensuring a safe passage.