Symphony: Learning Realistic and Diverse Agents for Autonomous Driving Simulation (2205.03195v1)

Abstract: Simulation is a crucial tool for accelerating the development of autonomous vehicles. Making simulation realistic requires models of the human road users who interact with such cars. Such models can be obtained by applying learning from demonstration (LfD) to trajectories observed by cars already on the road. However, existing LfD methods are typically insufficient, yielding policies that frequently collide or drive off the road. To address this problem, we propose Symphony, which greatly improves realism by combining conventional policies with a parallel beam search. The beam search refines these policies on the fly by pruning branches that are unfavourably evaluated by a discriminator. However, it can also harm diversity, i.e., how well the agents cover the entire distribution of realistic behaviour, as pruning can encourage mode collapse. Symphony addresses this issue with a hierarchical approach, factoring agent behaviour into goal generation and goal conditioning. The use of such goals ensures that agent diversity neither disappears during adversarial training nor is pruned away by the beam search. Experiments on both proprietary and open Waymo datasets confirm that Symphony agents learn more realistic and diverse behaviour than several baselines.

• The paper introduces a novel framework that leverages hierarchical policies and parallel beam search to improve simulation realism for autonomous driving.
• It utilizes discriminator-based adversarial training to iteratively refine agent trajectories, reducing collision and off-road events compared to traditional LfD methods.
• Experimental results on Waymo datasets show significant improvements in realism and diversity metrics, enhancing simulation fidelity for safe vehicle testing.

Symphony: Advancements in Autonomous Driving Simulation

The paper entitled "Symphony: Learning Realistic and Diverse Agents for Autonomous Driving Simulation" addresses a critical feature in the development of autonomous vehicles: the need for realistic and diverse simulations of human road users. Autonomous driving systems require extensive testing in varied scenarios to ensure safety and reliability. Real-world testing, however, can be risky and expensive, making high-fidelity simulation an attractive alternative.

Simulation Challenges and Learning from Demonstration (LfD)

To model human road users accurately, the paper leverages Learning from Demonstration (LfD), utilizing trajectories from sensors like LIDAR and cameras mounted on vehicles. Traditional LfD methods, such as behavioral cloning and generative adversarial imitation learning, often fall short by generating policies that, although optimized for loss functions, still result in unrealistic behavior like collisions or vehicles leaving the road.

Symphony Methodology

Symphony introduces a novel framework aimed at addressing these shortcomings. It enhances the realism of simulations by integrating a hierarchical policy structure with parallel beam search mechanisms.

1. Hierarchical Policies: Symphony's hierarchical approach segregates behavior into goal generation and goal conditioning, improving diversity and aligning with high-level intents.
2. Parallel Beam Search: By applying a beam search approach during simulations, unrealistic branches are pruned based on discriminator evaluations, focusing computations on the most promising outcomes. This real-time refinement during rollouts ensures realistic policy outputs and compels the training of more realistic models.
3. Discriminator Utilization: A discriminator continuously evaluates agent trajectory realism against actual data, helping to iteratively refine the agent policies through adversarial training.

Evaluations and Experimental Results

The paper conducts rigorous experiments using proprietary Waymo data and the open Waymo Open Motion Dataset to validate Symphony's efficacy. The results demonstrate Symphony's substantial improvements in generating more realistic behaviors compared to traditional methods such as BC and MGAIL. Importantly, metrics like collision rate and off-road time are markedly reduced, signifying enhanced realism.

The addition of hierarchical policies particularly addresses the issue of mode collapse observed when using beam search alone. While beam search enhances realism, it tends to result in mode collapse, compromising diversity. Hierarchical policies counteract this by ensuring the diverse representation of possible behaviors.

Furthermore, Symphony's agents outperform baselines in terms of both realism metrics (collision rate, off-road time, and ADE) and diversity measures (minSADE and curvature Jensen-Shannon divergence). The curvature JSD is a novel measure proposed by the authors, which reflects how well the high-level behaviors match observed empirical distributions.

Broader Implications and Future Work

Symphony offers significant implications for autonomous vehicle testing and development. By improving simulation realism, it facilitates more reliable and comprehensive testing scenarios without the safety risks associated with real-world trials. The balancing of realism and diversity enriches the synthetic data ecosystem, allowing developers to better prepare autonomous systems for complex, variable driving environments.

Future research may explore further refinement of the pruning mechanisms within the beam search or integration with models representing driver personas for personalizing agent behavior. Additional diversity metrics may also be developed to capture other aspects of agents’ behavior patterns, thereby enhancing Symphony's robustness in diverse environments.

In summary, Symphony represents a meaningful advance in simulation methodologies for autonomous driving, enhancing both the realism of simulated environments and the diversity of behaviors these simulations can explore. This work broadens the horizon of safe and effective testing for autonomous systems, paving the way for further innovations in AI-driven vehicular technologies.