Safe planning and control under uncertainty for self-driving (2010.11063v1)

Abstract: Motion Planning under uncertainty is critical for safe self-driving. In this paper, we propose a unified obstacle avoidance framework that deals with 1) uncertainty in ego-vehicle motion; and 2) prediction uncertainty of dynamic obstacles from environment. A two-stage traffic participant trajectory predictor comprising short-term and long-term prediction is used in the planning layer to generate safe but not over-conservative trajectories for the ego vehicle. The prediction module cooperates well with existing planning approaches. Our work showcases its effectiveness in a Frenet frame planner. A robust controller using tube MPC guarantees safe execution of the trajectory in the presence of state noise and dynamic model uncertainty. A Gaussian process regression model is used for online identification of the uncertainty's bound. We demonstrate effectiveness, safety, and real-time performance of our framework in the CARLA simulator.

• The paper presents a dual-layer trajectory prediction model that balances safety and efficiency by combining immediate reachability analysis with longer-term motion goals.
• It employs a robust tube MPC enhanced with Gaussian Process Regression to maintain safe trajectories amid control and state uncertainties.
• Simulations in the CARLA simulator validate the approach’s low computational cost and effective handling of complex driving scenarios.

Safe Planning and Control Under Uncertainty for Self-Driving

The discussed paper presents a comprehensive framework aimed at improving the safety and efficiency of self-driving vehicles when dealing with uncertainties inherent in real-world driving environments. The focus is on developing a system that accounts for both the prediction uncertainty of dynamic obstacles and the uncertainties related to the ego-vehicle’s control commands and state noise.

Unified Framework for Obstacle Avoidance

The researchers propose a unified approach that seamlessly integrates prediction, planning, and control, built around two primary innovations: a dual-layer trajectory prediction model and a robust control framework using tube-based Model Predictive Control (MPC).

1. Dual-layer Trajectory Prediction Model: This model employs a short-term prediction layer that utilizes reachability analysis to handle immediate uncertainties in dynamic obstacle trajectories. Complementing this, a long-term prediction layer is designed to factor in the larger-scale motion goals of these obstacles. The integration of these two layers aims to provide a balance between safety and efficiency, yielding trajectories that are both collision-free and non-conservative.
2. Robust Tube MPC: For control at the execution level, the researchers employ tube MPC, which ensures that the vehicle remains within a safe trajectory “tube” despite uncertainties. This approach also leverages a Gaussian Process Regression (GPR) model for real-time estimation of the uncertainty bounds, thus avoiding the impracticality of predefined error margins.

By utilizing the Frenet frame for trajectory planning, a method known for its efficacy in handling variable road geometries, the framework effectively generates and evaluates multiple potential paths, selecting the one with the optimal cost under the given constraints.

Numerical Results and Implications

The framework's efficacy is validated through a range of simulations in the high-fidelity CARLA simulator, capturing scenarios like vehicle overtaking, intersection navigation, and curved road traversal. These simulations, conducted under conditions of bounded state and control disturbances, demonstrate the framework's robust performance, with the ego vehicle maintaining safe trajectories amid unpredictable obstacles.

From a numerical standpoint, the advancements in this paper notably reduce computational requirements without sacrificing safety, making the approach suitable for real-time implementation in modern self-driving systems.

Practical and Theoretical Implications

The practical implications of this work are significant. By addressing both prediction and control-level uncertainties within a unified framework, the paper contributes towards enhancing the safety reliability of self-driving vehicles, an essential factor for widespread real-world deployment.

Theoretically, this research pushes the boundary of existing MPC applications by integrating advanced prediction models and uncertainty estimation techniques. The novel use of GPR for online uncertainty bounds presents a promising direction for future research, potentially applicable to various domains where real-time adaptability is required.

Future development might explore enhancing the prediction model using more sophisticated machine learning approaches, or expanding the robust control framework to accommodate a broader range of disturbance scenarios. Additionally, further work could aim to refine computational efficiency, a critical factor for the viability of real-time autonomous systems.

In conclusion, the paper delivers a meticulously detailed and effectively validated framework that strengthens both the theoretical foundation and practical application of autonomous vehicle control under uncertainty, marking a noteworthy contribution to the field of autonomous driving technology.