[论文解读] Data Driven Computational Model for Bipedal Walking and Push Recovery

Vijay Bhaskar Semwal|arXiv (Cornell University)|Jan 1, 2017

Gait Recognition and Analysis参考文献 64被引用 21

一句话总结

本文提出了一种基于数据驱动的计算模型，用于双足行走与推力恢复，采用混合自动机与元胞自动机，从人类运动数据生成关节轨迹。该模型通过分层类型-1模糊逻辑控制器实现稳定步态与有效的推力恢复，基于模拟与真实人类数据验证，轨迹预测与受力响应均具有高精度。

ABSTRACT

In this research, we have developed the data driven computational walking model to overcome the problem with traditional kinematics based model. Our model is adaptable and can adjust the parameter morphological similar to human. The human walk is a combination of different discrete sub-phases with their continuous dynamics. Any system which exhibits the discrete switching logic and continuous dynamics can be represented using a hybrid system. In this research, the bipedal locomotion is analyzed which is important for understanding the stability and to negotiate with the external perturbations. We have also studied the other important behavior push recovery. The Push recovery is also a very important behavior acquired by human with continuous interaction with environment. The researchers are trying to develop robots that must have the capability of push recovery to safely maneuver in a dynamic environment. The push is a very commonly experienced phenomenon in cluttered environment. The human beings can recover from external push up to a certain extent using different strategies of hip, knee and ankle. The different human beings have different push recovery capabilities. For example a wrestler has a better push negotiation capability compared to normal human beings. The push negotiation capability acquired by human, therefore, is based on learning but the learning mechanism is still unknown to researchers. The research community across the world is trying to develop various humanoid models to solve this mystery. Seeing all the conventional mechanics and control based models have some inherent limitations, a learning based computational model has been developed to address effectively this issue. In this research we will discuss how we have framed this problem as hybrid system.

研究动机与目标

• 开发一种基于数据驱动的双足步行计算模型，以克服传统基于运动学模型的局限性。
• 利用混合系统与元胞自动机对人类步态与推力恢复进行建模，实现状态预测与轨迹生成。
• 设计一种计算高效、分层的模糊逻辑控制器，用于推力恢复，可适应不同大小与方向的外力。
• 利用通过自定义可穿戴设备捕获的真实人类运动数据验证该模型，并与模拟与机器人平台进行对比。

提出的方法

• 开发了包含七个子阶段的混合自动机模型，用于表示双足步行中离散切换与连续动力学的结合。
• 利用通过HMCD与HLPRDCD设备捕获的人类运动数据导出的向量场，生成六个关节（髋、膝、踝）的关节轨迹。
• 应用具有16条规则的元胞自动机，通过4位流表示建模步态状态转换，实现实时状态预测且误差极小。
• 利用深度神经网络对推力恢复数据进行分类，并与其他机器学习技术进行对比，以验证鲁棒性。
• 设计分层类型-1模糊逻辑控制器，以降低计算负载，同时实现对外部扰动的快速、泛化性响应。
• 通过OpenSim gait 2354与HOAP2机器人模拟验证模型，关节轨迹输出与真实人类数据进行对比测试。

实验结果

研究问题

• RQ1 如何通过混合系统建模，利用数据驱动的计算模型准确复现人类双足步行的动力学？
• RQ2 何种方法可在保持精度的前提下，以最低计算成本建模步态状态转换？
• RQ3 如何构建模糊逻辑控制器，以实现在计算开销极小情况下的快速、泛化性推力恢复？
• RQ4 从人类数据生成的关节轨迹在形态相似的机器人平台上可实现多大程度的迁移？
• RQ5 不同大小与方向的外力如何影响关节角度响应？该模型能否准确预测这些响应？

主要发现

• 基于数据驱动的混合自动机模型成功生成了七个步态子阶段中所有六个关节（髋、膝、踝）的关节轨迹，与人类运动数据高度一致。
• 元胞自动机模型通过16条规则与4位流表示，实现了高精度的步态状态预测，支持实时状态跟踪且误差在可接受范围内。
• 分层类型-1模糊逻辑控制器表现出快速适应性与低计算成本，能有效预测不同外力大小（0–12 N）下的合适恢复策略。
• 该模型的关节轨迹成功应用于HOAP2机器人，证明了其在真实世界应用中的可行性。
• 与OpenSim gait 2354的对比验证确认了生成轨迹在不同运动阶段的一致性与高精度。
• 模糊控制器在不同受试者与力方向间具有良好泛化能力，特定关节角度范围（如大外力下髋关节：14.1°至-0.9°）与实际人类响应相符。