Code and data underlying the PhD thesis: Safe yet Precise Soft Robots via Incorporating Physics into Learned Models for Control

Code and data associated with the Ph.D. thesis titled "Safe yet Precise Soft Robots: Incorporating Physics into Learned Models for Control" by Maximilian Stölzle at the Department of Cognitive Robotics, Faculty of Mechanical Engineering, Delft University of Technology. For each (published) chapter, we release the code and the data underlying the figures and plots in the chapter as ZIP archives.<strong>chapter_4: </strong>MATLAB code for running ORB-SLAM and subsequently projecting the pose estimates into the soft robot kinematic model. Additionally, it contains the configuration files for a Blender environment that was used for validation in simulation.<strong>chapter_5:</strong> We include the HSA PyElastica package for simulating HSA robots (<em>hsa-actuation-matlab</em>) and the SPCS kinematic model (<em>jax-spcs-kinematics</em>) implemented in JAX for capturing the deformation of HSA rods. The <em>hsa-kinematic-model</em> allows for regressing the configuration of the HSA rod by running differential inverse kinematics. It also contains MATLAB code for actuating the HSA robot and collecting datasets. The <em>jax-soft-robot-modeling</em> package contains a reduced-order, control-oriented kinematic, and dynamic model for planar HSA robots, which subsequently allows simulation.<strong>chapter_6:</strong> It contains the data and the code for identifying the system parameters of planar HSA robots based on the dynamic model presented in Chapter 5. Furthermore, it contains ROS2 packages for running closed-loop control experiments with the HSA robot in the planar setting. For this purpose, we implement model-based controllers in the <em>hsa-planar-control</em> package, which also contains the experimental data and scripts for analyzing and plotting the results.<strong>chapter_7:</strong> It contains the code and data for the guiding of planar HSA robots using brain signals. For this purpose, we include again the JAX soft robot modeling package, the ROS2 packages for operating HSA robots, and the model-based planar HSA controllers. Additionally, we embed the <em>sr-brain-control</em> package that contains the experimental data and the scripts for plotting them. Additionally, it contains the OpenVibe specifications of the EEG preprocessing and classification pipeline.<strong>chapter_8: </strong>This archive includes the data and MATLAB code for the backstepping controller and the associated closed-loop simulations.<strong>chapter_9:</strong> We include the Python <em>promasens</em> package that includes the code for soft robot shape sensing based on magnetic sensors and the associated data to train the neural networks and plot the simulation and experimental results. Furthermore, we included a ROS2 package for feedforward actuation sequences for pneumatic soft robots that we used to collect the datasets.<strong>chapter_10:</strong> We include the Python code to perform kinematic fusion, dynamic identification, and the scripts and data for generating the plots included in the chapter.<strong>chapter_11:</strong> We include the code for training latent dynamics using Coupled Oscillator Networks and subsequently leveraging the learned dynamics for model-based control in latent space.<strong>appendix_c: </strong>We re-include the most important software packages developed as part of this Ph.D that are motivated, introduced, and explained in Appendix C of this thesis. First, it contains the JAX Soft Robot Modeling package and the model-based controllers for planar HSA robots. This includes the ROS2 packages for communicating with the Festo VTEM pressure regulator (<em>ros2-vtem_control</em>), the Optitrack motion capture system (ros2-mocap_optitrack), and operating the HSA robot. Furthermore, we include a ROS2 package that generates pressure trajectories for the actuation of pneumatic soft robots moving in 3D space (<em>ros2-pneumatic_actuation</em>). Finally, we include a package that builds Docker containers for operating (HSA) soft robots that enable rapid bootstrapping of a ROS2 environment on a new workstation (<em>sr-ros2-bundles</em>). The docker images are automatically built nightly via GitHub Actions CI.<br>

本数据集配套代码与数据来自代尔夫特理工大学机械工程学院认知机器人系Maximilian Stölzle的博士论文《Safe yet Precise Soft Robots: Incorporating Physics into Learned Models for Control》（安全且精准的软体机器人：将物理特性融入学习型控制模型）。针对每一篇已发表的章节，我们均以ZIP压缩包形式发布该章节中图表所依托的代码与数据。<strong>第4章：</strong>包含用于运行ORB-SLAM，并将位姿估计结果投影至软体机器人运动学模型的MATLAB代码。此外还包含用于仿真验证的Blender环境配置文件。<strong>第5章：</strong>提供了用于仿真形状记忆合金（HSA）机器人的PyElastica HSA工具包<em>hsa-actuation-matlab</em>，以及基于JAX实现的、用于捕捉HSA杆形变的SPCS运动学模型<em>jax-spcs-kinematics</em>。<em>hsa-kinematic-model</em>工具包可通过差分逆运动学回归得到HSA杆的构型，同时包含用于驱动HSA机器人与采集数据集的MATLAB代码。<em>jax-soft-robot-modeling</em>工具包则提供了面向平面HSA机器人的降阶、面向控制的运动学与动力学模型，可用于仿真。<strong>第6章：</strong>包含用于基于第5章提出的动力学模型辨识平面HSA机器人系统参数的代码与数据。此外还包含用于在平面场景下开展HSA机器人闭环控制实验的ROS2工具包。其中，<em>hsa-planar-control</em>工具包实现了基于模型的控制器，同时包含实验数据以及用于分析与绘图的脚本。<strong>第7章：</strong>包含用于利用脑信号引导平面HSA机器人的代码与数据。其中再次提供了JAX软体机器人建模工具包、用于操作HSA机器人的ROS2工具包，以及基于模型的平面HSA控制器。此外还包含<em>sr-brain-control</em>工具包，该工具包提供了实验数据与绘图脚本。同时还包含脑电图（EEG）预处理与分类流水线的OpenVibe规范文件。<strong>第8章：</strong>该压缩包包含用于反步控制器及其相关闭环仿真的实验数据与MATLAB代码。<strong>第9章：</strong>提供了Python工具包<em>promasens</em>，其中包含基于磁传感器的软体机器人形状感知代码，以及用于训练神经网络、绘制仿真与实验结果的配套数据。此外还包含一个ROS2工具包，用于为气动软体机器人生成前馈驱动序列，该工具包曾用于数据集采集。<strong>第10章：</strong>提供了用于开展运动学融合与动力学辨识的Python代码，以及用于生成该章节图表的脚本与数据。<strong>第11章：</strong>提供了用于基于耦合振子网络（Coupled Oscillator Networks）训练隐式动力学，并将学习得到的动力学用于隐空间模型控制的代码。<strong>附录C：</strong>重新收录了本博士论文附录C中所介绍、阐释的核心软件开发包。首先包含JAX软体机器人建模工具包，以及用于平面HSA机器人的基于模型的控制器。其中包含用于与费斯托（Festo）VTEM压力调节器通信的ROS2工具包<em>ros2-vtem_control</em>、用于Optitrack动作捕捉系统的ROS2工具包ros2-mocap_optitrack，以及用于操作HSA机器人的工具包。此外还包含一个用于为在三维空间运动的气动软体机器人生成驱动压力轨迹的ROS2工具包<em>ros2-pneumatic_actuation</em>。最后还提供了一个用于构建（HSA）软体机器人操作系统的Docker容器工具包<em>sr-ros2-bundles</em>，可用于在新工作站上快速搭建ROS2环境。该Docker镜像通过GitHub Actions CI实现每日自动构建。