Enhanced Safety in Autonomous Driving: Integrating Latent State Diffusion Model for End-to-End Navigation

TL;DR

The paper addresses safety in end-to-end autonomous driving by formulating navigation as a safety-constrained CMDP and introducing ESAD-LEND, which combines latent diffusion-based state representation, a worst-case exploration strategy, and a CVaR-enhanced Soft Actor-Critic within an augmented Lagrangian framework. It enables latent imagination of future trajectories and safety-guaranteed planning, backed by distributional RL to handle risk via CVaR. Evaluations in CARLA and real-world-like settings show ESAD-LEND achieving superior safety, efficiency, and generalization compared with several baselines. This work advances practical safe exploration in high-dimensional autonomous driving and offers a principled, model-based approach for end-to-end navigation with rigorous safety guarantees.

Abstract

With the advancement of autonomous driving, ensuring safety during motion planning and navigation is becoming more and more important. However, most end-to-end planning methods suffer from a lack of safety. This research addresses the safety issue in the control optimization problem of autonomous driving, formulated as Constrained Markov Decision Processes (CMDPs). We propose a novel, model-based approach for policy optimization, utilizing a conditional Value-at-Risk based Soft Actor Critic to manage constraints in complex, high-dimensional state spaces effectively. Our method introduces a worst-case actor to guide safe exploration, ensuring rigorous adherence to safety requirements even in unpredictable scenarios. The policy optimization employs the Augmented Lagrangian method and leverages latent diffusion models to predict and simulate future trajectories. This dual approach not only aids in navigating environments safely but also refines the policy's performance by integrating distribution modeling to account for environmental uncertainties. Empirical evaluations conducted in both simulated and real environment demonstrate that our approach outperforms existing methods in terms of safety, efficiency, and decision-making capabilities.

Figures (9)

• Figure 1: An overview of our proposed framework. Enhanced safe navigation Autonomous Driving model with Latent State End-to-Navigation Diffusion model (ESAD-LEND) The process initiates as the agent takes an action in the environment and subsequently receives the state $S$ and reward $r$. The state is encoded into a latent space $Z$ using an encoder. This latent representation is then decoded to simulate environmental states for decision-making. The policy $\pi(a|s, Z)$ leverages these latent representations to determine appropriate actions $a_1, a_2, \ldots, a_n$ across states $S_1, S_2, \ldots, S_n$.
• Figure 2: The policy generation process in a diffusion-based control system. Starting with a state representation of the environment, the diffusion model $P(z|s)$ generates multiple candidate latent states. These candidates are evaluated by a Q-function $\arg\max Q(s, z)$, selecting the optimal latent state for action execution. The process integrates a safety guarantee mechanism to ensure that the chosen action meets predefined safety criteria before execution, culminating in a controlled action that adheres to safety standards.
• Figure 3: The policy generation process in a diffusion-based control system. Starting with a state representation of the environment, the diffusion model $P(z|s)$ generates multiple candidate latent states. These candidates are evaluated by a Q-function $\arg\max Q(s, z)$, selecting the optimal latent state for action execution. The process integrates a safety guarantee mechanism to ensure that the chosen action meets predefined safety criteria before execution, culminating in a controlled action that adheres to safety standards.
• Figure 4: Schematic representation of the world model learning process in our autonomous system. The process begins with the state input into the diffusion model $P(z|s)$, which generates multiple candidate latent states. These states are evaluated using a Q-function, $\arg\max Q(s, z)$, to select the optimal state for action execution, ensuring that it meets safety criteria. The selected action is then applied in the world model learning component, which simulates possible future trajectories based on continuous actor-critic policy optimization and updates in response to new data from the environment. This iterative process enhances the agent’s ability to predict and react to dynamic scenarios, integrating safety constraints at every step.
• Figure 5: Illustration of various safety-critical scenarios developed to assess the response capabilities of autonomous driving systems. These scenarios include Traffic Negotiation, where multiple vehicles interact at an intersection; Highway, depicting lane changes and merges under high-speed conditions; Obstacle Avoidance, showing responses to unexpected roadblocks; and Braking and Lane Changing, which involves rapid deceleration and maneuvering to avoid collisions. These tests are crucial for validating the robustness and reliability of safety protocols in autonomous vehicles.