登录 EN

添加临时用户

基于生成模型的安全轨迹规划

Generative Model-Based Safe Trajectory Planning

作者:黄燚
  • 学号
    2022******
  • 学位
    硕士
  • 电子邮箱
    110******com
  • 答辩日期
    2025.05.14
  • 导师
    刘厚德
  • 学科名
    电子信息
  • 页码
    70
  • 保密级别
    公开
  • 培养单位
    599 国际研究生院
  • 中文关键词
    扩散模型;模型预测路径积分;机器人规划;未知环境;条件流匹配
  • 英文关键词
    Diffusion Model; Model Predictive Path Integral; Robot Planning? Unknown Environment? Conditional Flow Matching

摘要

在复杂和未知环境中实现快速、安全、可靠的机械臂轨迹规划是机器人领域的核心挑战。传统的轨迹规划方法通常难以同时满足速度、完备性和最优性等多重需求,或依赖于精确的环境模型,难以应对复杂和未知场景。而近年来兴起的基于生成模型的规划方法和基于模型预测路径积分 (MPPI) 的方法虽然展现出了强大的潜力,但前者通常缺乏安全保障,后者则由于其局部特性而易于陷入局部最优。针对上述问题,本文提出了一种基于生成模型引导的模型预测路径积分安全运动生成框架,该框架主要用于解决复杂和未知环境下的安全运动生成问题,包含以下研究内容:首先,本文提出了一种基于扩散模型引导 MPPI 采样的运动生成方法。该方法设计了一种生成加优化的两阶段规划框架,通过采用扩散模型生成 MPPI 算法控制序列均值的方式,有效引导其在复杂环境中进行更高效的探索和采样,避免陷入局部最优。针对扩散模型与 MPPI 模块工作频率不一致的问题,本文提出了混合采样策略和异步更新策略,降低了扩散模型生成速度慢带来的不利影响,实现了全局引导与局部探索的平衡。通过在复杂 2 维和 3 维环境中的实验,验证了所提出方法的有效性。其次,针对上述方法在未知环境中泛化能力不足的问题,本文扩展了上述基本框架,提出将环境点云信息作为条件向量嵌入到控制序列均值的生成过程中,引导扩散模型采样过程生成与环境上下文相适应的全局轨迹。此外,引入了基于点云的碰撞检测方法,用于设计 MPPI 算法的成本函数,进一步拓展了框架的泛化能力。实验结果表明改进后的方法能够适应训练中未见过的环境,仍然能保持优秀的轨迹生成能力和任务成功率。最后,针对扩散模型计算效率低下的问题,本文提出了基于条件流匹配的改进方法,采用线性插值路径作为概率传输路径,将轨迹生成问题建模为学习从随机噪声分布到目标轨迹分布的连续概率流。通过神经网络学习数据空间中引导数据点运动的速度场,驱动噪声分布向目标分布变换。在保证未知环境中任务成功率的前提下,基于条件流匹配的方法实现了比基于扩散模型方法近 10 倍的轨迹生成速度,提升了算法的实时性。

Achieving fast, safe, and reliable trajectory planning for robotic manipulators in complex and unknown environments presents a key challenge in the field of robotics. Traditional trajectory planning methods often struggle to simultaneously satisfy multiple requirements such as speed, completeness, and optimality. They also tend to rely on precise environmental models, making them less suitable for complex and unknown scenarios. While recently emerging methods based on generative models and Model Predictive Path Integral (MPPI) show considerable potential, the former often lack safety guarantees, and the latter are prone to local optima due to their inherent local search characteristics. To address these issues, this paper proposes a safe trajectory planning framework guided by a generative model for Model Predictive Path Integral. This framework primarily aims to solve safe trajectory planning problems in complex and unknown environments and encompasses the following research contents:First, this paper introduces a trajectory planning method based on diffusion model-guided MPPI sampling. This method employs a two-stage “generate-then-optimize” planning framework. By using a diffusion model to generate the mean of the MPPI algorithm’s control sequences, it effectively guides MPPI to explore and sample more efficiently in complex environments, thereby avoiding local optima. To address the inconsistent operating frequencies between the diffusion model and the MPPI module, this paper proposes a hybrid sampling strategy and an asynchronous update strategy. These strategies reduce the negative impacts of the diffusion model’s slow generation speed, achieving a balance between global guidance and local exploration. The effectiveness of the proposed method is validated through experiments in complex 2D and 3D environments.Second, to address the limited generalization ability of the aforementioned method in unknown environments, this paper extends the basic framework. It proposes integrating environmental point cloud information as a conditional input during the generation of the control sequence mean. This guides the diffusion model’s sampling process to produce global trajectories adapted to the environmental context. Furthermore, a point cloud-based collision detection method is introduced for designing the MPPI algorithm’s cost function, further improving the framework’s ability to generalize. Experimental results demonstrate that the improved method can adapt to environments not encountered during training, while still maintaining excellent trajectory generation capabilities and task success rates.Finally, to address the low computational efficiency of diffusion models, this paper proposes an improved method based on Conditional Flow Matching (CFM). It uses linear interpolation paths for probability transport and models trajectory generation as learning a continuous probability flow from a random noise distribution to a target trajectory distribution. A neural network learns the velocity field in the data space that directs the motion of data points, transforming the noise distribution towards the target distribution. While ensuring task success rates in unknown environments, the CFM-based method achieves trajectory generation speeds nearly 10 times faster than the diffusion model-based method, thereby improving the algorithm’s real-time capability. The effectiveness of this proposed method is demonstrated in both static and dynamic environments.