Autonomous Navigation of AGVs in Unknown Cluttered Environments: log-MPPI Control Strategy (2203.16599v2)

Abstract: Sampling-based model predictive control (MPC) optimization methods, such as Model Predictive Path Integral (MPPI), have recently shown promising results in various robotic tasks. However, it might produce an infeasible trajectory when the distributions of all sampled trajectories are concentrated within high-cost even infeasible regions. In this study, we propose a new method called log-MPPI equipped with a more effective trajectory sampling distribution policy which significantly improves the trajectory feasibility in terms of satisfying system constraints. The key point is to draw the trajectory samples from the normal log-normal (NLN) mixture distribution, rather than from Gaussian distribution. Furthermore, this work presents a method for collision-free navigation in unknown cluttered environments by incorporating the 2D occupancy grid map into the optimization problem of the sampling-based MPC algorithm. We first validate the efficiency and robustness of our proposed control strategy through extensive simulations of 2D autonomous navigation in different types of cluttered environments as well as the cartpole swing-up task. We further demonstrate, through real-world experiments, the applicability of log-MPPI for performing a 2D grid-based collision-free navigation in an unknown cluttered environment, showing its superiority to be utilized with the local costmap without adding additional complexity to the optimization problem. A video demonstrating the real-world and simulation results is available at https://youtu.be/_uGWQEFJSN0.

• The paper introduces a log-MPPI strategy that employs a novel NLN mixture distribution for trajectory sampling to maintain feasibility under complex constraints.
• It demonstrates improved AGV navigation performance with faster convergence and higher success rates in both simulated and real-world cluttered scenarios.
• Results highlight the method’s efficiency in reducing computational overhead and constraint violations compared to traditional MPPI techniques.

Overview of "Autonomous Navigation of AGVs in Unknown Cluttered Environments: log-MPPI Control Strategy"

This paper presents an enhanced control strategy for autonomous ground vehicles (AGVs) navigating in unknown, cluttered environments. The proposed "log-MPPI" control strategy builds upon the traditional Model Predictive Path Integral (MPPI) framework by integrating a novel trajectory sampling approach, which improves trajectory feasibility with respect to system constraints.

The log-MPPI method innovatively utilizes trajectory sampling from a normal log-normal (NLN) mixture distribution, diverging from the conventional Gaussian distribution used in previous MPPI applications. This transition to an NLN-based sampling mechanism is central to the presented algorithm's ability to enhance the feasibility of sampled trajectories, especially in non-convex and constraint-dense environments, a characteristic challenge in robotic navigation.

Key Contributions and Methodology

The proposed log-MPPI strategy focuses on sampling efficiency and robustness in constraint satisfaction. A noteworthy aspect of this approach is the use of NLN mixture distribution for sampling trajectories, which allows maintaining a lower noise variance, thereby minimizing constraint violations. This addresses the inherent issue with classical sampling methods which tend to congregate in high-cost or non-feasible regions, potentially leading to infeasible trajectory outcomes.

In comparison to existing methods which may require additional auxiliary controllers or feedback mechanisms to ensure system constraint satisfaction, the log-MPPI approach minimizes additional computational overhead. It achieves this through careful distribution selection, yielding desirable state-space exploration while reducing the complexity associated with traditional optimization constraints.

Results and Validation

The paper substantiates the efficacy of the log-MPPI strategy through extensive simulations and real-world experiments. Simulations in dynamic and static cluttered environments alongside a benchmark cartpole swing-up task highlight the robustness and adaptability of the presented method. Results indicate that log-MPPI consistently outperforms conventional MPPI under similar or lower noise variance conditions, achieving higher success rates and quicker convergence to feasible trajectories. This is particularly evident in environments where obstacle density varies appreciably.

The empirical results emphasize the control strategy's ability to successfully navigate complex spaces while adhering to constraints without inundating the computational architecture, benefiting from parallelism inherent in compute platforms such as GPUs. Additionally, real-world experiments demonstrate the practical applicability of the log-MPPI in executing real-time navigation tasks using occupancy grid maps for environmental perception.

Implications and Future Outlook

The implication of this research extends to broader applications in robotics and autonomous systems where safe, efficient navigation in unpredictable environments is crucial. By providing a mechanism for enhanced exploration and reduced risk of suboptimal trajectory selection, log-MPPI elevates the reliability of autonomous navigation systems.

Future developments may focus on scaling the proposed method to operate efficiently on standard computational hardware, thereby broadening its applicability in cost-sensitive deployments. Further research could explore the integration of adaptive sensing and dynamic feedback mechanisms to enhance system robustness against environmental uncertainties and sensor noise. The theoretical exploration of stability and convergence guarantees for sampling-based MPC frameworks under NLN distributions remains a potentially rich area of investigation.