Abstract:Inverse dynamics (ID), which aims at reproducing the driven torques from human kinematic observations, has been a critical tool for gait analysis. However, it is hindered from wider application to general motion due to its limited scalability. Conventional optimization-based ID requires expensive laboratory setups, restricting its availability. To alleviate this problem, we propose to exploit the recently progressive human motion imitation algorithms to learn human inverse dynamics in a data-driven manner. The key insight is that the human ID knowledge is implicitly possessed by motion imitators, though not directly applicable. In light of this, we devise an efficient data collection pipeline with state-of-the-art motion imitation algorithms and physics simulators, resulting in a large-scale human inverse dynamics benchmark as Imitated Dynamics (ImDy). ImDy contains over 150 hours of motion with joint torque and full-body ground reaction force data. With ImDy, we train a data-driven human inverse dynamics solver ImDyS(olver) in a fully supervised manner, which conducts ID and ground reaction force estimation simultaneously. Experiments on ImDy and real-world data demonstrate the impressive competency of ImDyS in human inverse dynamics and ground reaction force estimation. Moreover, the potential of ImDy(-S) as a fundamental motion analysis tool is exhibited with downstream applications. The project page is <a class="link-external link-https" href="https://foruck.github.io/ImDy/" rel="external noopener nofollow">this https URL</a>.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: **How to accurately estimate human inverse dynamics (ID), including driving torque and ground reaction force (GRF), from pure kinematic data?** Traditional methods rely on expensive laboratory equipment and are difficult to extend to general motion scenarios. Therefore, the paper proposes a data - driven method based on imitation learning (IL) to solve this problem. ### Detailed Explanation 1. **Background and Challenges**: - **Traditional Inverse Dynamics Methods**: Traditional human inverse dynamics methods mainly calculate driving torque through optimization algorithms, which require accurate ground reaction force measurements. This usually depends on expensive laboratory equipment and is very sensitive to small motion - capture noises. - **Scalability Issues**: These methods are difficult to extend to broader general motion scenarios in practical applications, such as in the fields of robotics, healthcare, and sports training. 2. **Proposed Method**: - **The Relationship between Imitation Learning and Inverse Dynamics**: The author found that although imitation learning algorithms are mainly used to imitate the kinematic characteristics of human motion, they also implicitly contain an understanding of human inverse dynamics. - **Data Collection Pipeline**: To solve the bottleneck problem of data acquisition, the author designed an efficient data collection pipeline, using advanced physical simulators and imitation learning algorithms to generate a large - scale human inverse dynamics dataset ImDy. ImDy contains more than 150 hours of motion data, covering joint torque and whole - body ground reaction force. - **Data - Driven Inverse Dynamics Solver ImDyS**: Based on the ImDy dataset, the author trained a data - driven inverse dynamics solver ImDyS, which can simultaneously estimate driving torque and ground reaction force in a fully supervised manner. 3. **Experimental Verification**: - **Performance Evaluation**: Through extensive experiments on datasets such as ImDy, GroundLink, and AddBiomechanics, the effectiveness of ImDyS was verified. The experimental results show that ImDyS performs excellently in estimating driving torque and ground reaction force, especially having good generalization ability when dealing with real - world data. - **Comparative Analysis**: Compared with the traditional imitation learning algorithm PHC, ImDyS not only has a significant improvement in accuracy, but can also better handle the influence of kinematic errors on dynamic estimation. ### Summary The main contribution of this paper is to propose a large - scale human inverse dynamics data collection method based on imitation learning and develop a data - driven inverse dynamics solver ImDyS. This method not only solves the problem that traditional methods rely on expensive equipment, but also greatly expands the application range of inverse dynamics, providing new ideas and tools for future research.

ImDy: Human Inverse Dynamics from Imitated Observations

Off-Dynamics Inverse Reinforcement Learning

Dynamic Inertial Poser (DynaIP): Part-Based Motion Dynamics Learning for Enhanced Human Pose Estimation with Sparse Inertial Sensors

D&D: Learning Human Dynamics from Dynamic Camera

Homogeneous Dynamics Space for Heterogeneous Humans

Optimal-State Dynamics Estimation for Physics-based Human Motion Capture from Videos

PhysHOI: Physics-Based Imitation of Dynamic Human-Object Interaction

An Attention-Based Deep Learning Approach for Inertial Motion Recognition and Estimation in Human-Robot Collaboration

Toward Human Motion Digital Twin: A Motion Capture System for Human-Centric Applications

Open-source software library for real-time inertial measurement unit data-based inverse kinematics using OpenSim

DROP: Dynamics Responses from Human Motion Prior and Projective Dynamics

Physical Inertial Poser (PIP): Physics-aware Real-time Human Motion Tracking from Sparse Inertial Sensors

A forward-inverse dynamics modeling framework for human musculoskeletal multibody system

A complete process for human dynamics analysis including parameter identification and sEMG-torque estimation

Dynamics as Prompts: In-Context Learning for Sim-to-Real System Identifications

Transformer Inertial Poser: Real-time Human Motion Reconstruction from Sparse IMUs with Simultaneous Terrain Generation

Inverse dynamics with rigid contact and friction

Curriculum-based humanoid robot identification using large-scale human motion database

Forward Dynamics Estimation from Data-Driven Inverse Dynamics Learning

DiffusionPoser: Real-time Human Motion Reconstruction From Arbitrary Sparse Sensors Using Autoregressive Diffusion