Solving Motion Planning Tasks with a Scalable Generative Model (2407.02797v1)

Abstract: As autonomous driving systems being deployed to millions of vehicles, there is a pressing need of improving the system's scalability, safety and reducing the engineering cost. A realistic, scalable, and practical simulator of the driving world is highly desired. In this paper, we present an efficient solution based on generative models which learns the dynamics of the driving scenes. With this model, we can not only simulate the diverse futures of a given driving scenario but also generate a variety of driving scenarios conditioned on various prompts. Our innovative design allows the model to operate in both full-Autoregressive and partial-Autoregressive modes, significantly improving inference and training speed without sacrificing generative capability. This efficiency makes it ideal for being used as an online reactive environment for reinforcement learning, an evaluator for planning policies, and a high-fidelity simulator for testing. We evaluated our model against two real-world datasets: the Waymo motion dataset and the nuPlan dataset. On the simulation realism and scene generation benchmark, our model achieves the state-of-the-art performance. And in the planning benchmarks, our planner outperforms the prior arts. We conclude that the proposed generative model may serve as a foundation for a variety of motion planning tasks, including data generation, simulation, planning, and online training. Source code is public at https://github.com/HorizonRobotics/GUMP/

• The paper presents GUMP, a generative model that integrates full and partial autoregressive modes to improve training and inference speeds.
• It employs a convolutional encoder and a multimodal causal transformer with gated cross-attention to fuse static and dynamic traffic data.
• GUMP outperforms benchmarks by enhancing simulation realism, scenario generation, and interactive planning on Waymo and nuPlan datasets.

Solving Motion Planning Tasks with a Scalable Generative Model

The paper "Solving Motion Planning Tasks with a Scalable Generative Model" discusses the development of a Generative Unified Model for Motion Planning (GUMP) aimed at enhancing the scalability, safety, and cost-efficiency of autonomous driving systems. The authors position GUMP as a foundational model capable of supporting a range of motion planning tasks, demonstrating significant advancements in scenario generation, simulation realism, and planning capabilities.

Core Methodological Contributions

The central contribution of this paper is the proposal of GUMP, a scalable generative model integrating both full-autoregressive and partial-autoregressive modes. This dual-mode operation enhances both training and inference efficiencies, crucial for real-time applications in autonomous driving where computational resources and response times are pivotal.

1. Model Architecture: GUMP leverages a combination of convolutional encoder for static information and a multimodal causal transformer augmented with gated cross-attention blocks for dynamic and static information fusion. This innovative design empowers the model to capture complex traffic dynamics and agent interactions effectively.
2. Tokenization Approach: The research introduces a "key-value pair" tokenization strategy. This approach quantizes the state space with high granularity, facilitating structured and efficient state encoding. By adopting a key-value pair system, the model is capable of detailed and flexible manipulation of dynamic traffic scenarios, allowing for efficient management of agent appearance and disappearance.
3. Temporal Aggregation: To combat prediction errors inherent in autoregressive models, the authors develop a temporal aggregation strategy. This mechanism averages predictions over time, stabilizing output trajectories and enhancing simulation reliability.

Performance and Impact

The paper reports state-of-the-art results on several benchmarks including the Waymo Open Motion Dataset and the nuPlan planning dataset. GUMP outperforms existing solutions in key metrics:

• Simulation Realism: GUMP achieves high marks on the Waymo Sim Agents Benchmark, particularly in kinematics and interaction metrics, indicating improved realism in modeling agent dynamics and behaviors.
• Scene Generation: The model excels in scenario diversity and control, as evidenced by significant reductions in positional and velocity discrepancies compared to ground truth distributions.
• Interactive Planning: In the nuPlan dataset, GUMP showcases the effectiveness of its planning strategies, surpassing previous models in terms of driving scores and maintaining high compliance with traffic rules under varying conditions.

Implications and Future Work

GUMP's comprehensive framework suggests a paradigm shift in how autonomous systems might be continuously improved and evaluated. Its capability to generate realistic, interactive scenarios at scale has substantial implications for reducing reliance on real-world data collection, which is often cost-prohibitive and time-consuming. The model provides a promising platform for further research in diverse driving environments and conditions.

Future work might involve refining model accuracy through integration with vectorized map inputs or sensor data, enabling more nuanced scene understanding and prediction. Additionally, exploring the use of GUMP in multi-agent settings or complex vehicular negotiations could further enhance its applicability and robustness. The scalability, evidenced by improved performance with increased model capacity, aligns with trends in large model utilization, indicating fruitful avenues for research extension.

In summary, this paper introduces a novel and effective approach to autonomous vehicle motion planning, leveraging generative modeling to effectively simulate and plan in dynamic traffic environments. GUMP stands as a significant contribution to the development of scalable and efficient automated driving technologies.