Dual Control for Interactive Autonomous Merging with Model Predictive Diffusion

TL;DR

The paper addresses the challenge of interactive decision-making for autonomous merging under human-driven, uncertain behavior. It introduces a dual control framework that couples online Bayesian belief updates of human intents with a novel multimodal model-based diffusion solver designed for receding-horizon optimization. The approach is validated on real-time hardware using F1-Tenth platforms, showing that active belief probing enables earlier, safer, and more efficient merges compared to baselines. By integrating belief prediction and diffusion-based planning, the work advances interaction-aware planning and decision-making under uncertainty in autonomous driving, with practical implications for real-world highway merging scenarios.

Abstract

Interactive decision-making is essential in applications such as autonomous driving, where the agent must infer the behavior of nearby human drivers while planning in real-time. Traditional predict-then-act frameworks are often insufficient or inefficient because accurate inference of human behavior requires a continuous interaction rather than isolated prediction. To address this, we propose an active learning framework in which we rigorously derive predicted belief distributions. Additionally, we introduce a novel model-based diffusion solver tailored for online receding horizon control problems, demonstrated through a complex, non-convex highway merging scenario. Our approach extends previous high-fidelity dual control simulations to hardware experiments, which may be viewed at https://youtu.be/Q_JdZuopGL4, and verifies behavior inference in human-driven traffic scenarios, moving beyond idealized models. The results show improvements in adaptive planning under uncertainty, advancing the field of interactive decision-making for real-world applications.

Figures (6)

• Figure 1: Snapshots of traffic merge experiment at 8-second increments, when the ego vehicle must overtake (left) or yield (right) to merge. Passive learning with EMPPI (left), active learning with DMPPI (middle) active learning with proposed DMPD (right). A video of several experiments is available at https://youtu.be/Q_JdZuopGL4.
• Figure 2: Belief prediction approach enabling active learning.
• Figure 3: Model Predictive Diffusion: In the forward pass, $N_m$ samples from the previous time-step's optimal distribution $s^{0}(\cdot | x_{t})$ are corrupted with noise to form the dynamic prior distribution for the current time-step. In the backward pass, samples are drawn from the prior distribution $s^{N_d}(\cdot | x_{t+1})$ and are denoised to construct samples from the updated optimal distribution $s^{0}(\cdot | x_{t+1})$. The samples $u_{t:t+N-1}^{\tau, j}$ are indexed by their step in the diffusion process $\tau$, their sample index (also referred to as mode of the prior distribution which is a mixture of $N_d$ Gaussians) $j$, and their time-step in real time $t$.
• Figure 4: Merge scenario.
• Figure 5: F1-Tenth platform.

Theorems & Definitions (22)

• Lemma 1
• proof
• Theorem 1
• proof
• Corollary 1
• proof
• Proposition 1: knaup2024active
• Lemma 2
• proof
• Remark 1