Analyzing whole-body pose transitions in multi-contact motions

Christian Mandery, Júlia Borràs Sol, Mirjam Jöchner, Tamim Asfour

Abstract Humanoid robots often struggle with stability in complex environments where humans naturally utilize multi-contact strategies, such as leaning on surfaces or using handrails. This study analyzes over 1,300 human motion transitions from the KIT Whole-Body Human Motion Database to quantify how humans transition between different support poses involving hands, knees, and feet. By developing an automated contact detection algorithm and a transition graph, the researchers provide a statistical foundation for autonomous robot motion planning in multi-contact scenarios.

Introduction

While humanoid robots have achieved proficiency in bipedal locomotion on flat surfaces, they remain significantly less robust than humans in unpredictable or constrained environments. Humans instinctively maintain balance by reaching for handrails, leaning against walls, or utilizing their knees for support—behaviors categorized as multi-contact motion.

Despite theoretical models for such motions, empirical data on actual human execution is sparse. This research, conducted at the Karlsruhe Institute of Technology, addresses this gap by quantifying "loco-manipulation" patterns. The goal is to move beyond theoretical "sketch figures" toward autonomous, robust motion planning for humanoid robots.

Methodology

The researchers utilized a subset of the KIT Whole-Body Human Motion Database, consisting of 121 distinct motions recorded via a Vicon motion capture system. These motions were mapped onto the Master Motor Map (MMM), a unified reference model with 104 degrees of freedom.

Automated Contact Detection

A critical component of the study was the development of an automated algorithm to detect support contacts. A body segment (hand, foot, knee, or elbow) is classified as a "support" based on two criteria:

  1. Proximity: The segment must be within a specific distance threshold of an environmental object (e.g., 15mm for hands).
  2. Velocity: The segment’s speed must remain below $200 mm/s$ for a set duration, ensuring it is providing stable support rather than passing through the space.
Examples of multi-contact poses analyzed: (a) stair climbing with handrail, (b) leaning while picking up an object, (c) balancing on one leg with hand support, and (d) kneeling.

The resulting data was used to construct a Transition Graph, where nodes represent support poses (e.g., "2 Feet + 1 Hand") and edges represent the transitions between them. The dataset covered locomotion (stairs), loco-manipulation (leaning while picking up objects), balancing (recovery from pushes), and kneeling.

Results and Key Insights

The study identified and analyzed 1,323 individual pose transitions, totaling over nine minutes of high-resolution motion data. The automated segmentation method proved highly accurate, with a mere 2.13% error rate.

Bimodal Transition Dynamics

A significant finding was that transition durations are not uniform but often follow a bimodal distribution. For instance, transitions between one and two feet can be rapid (during walking) or prolonged (during deliberate manipulation). This suggests that robotic controllers must account for task context when determining transition timing.

The Role of "Light Touch"

The data supports the importance of "light touch"—contacts that provide sensory feedback to reduce body sway without necessarily bearing significant weight. This indicates that environmental support serves both mechanical and informational purposes.

Transition Complexity

While theoretical models often assume a "one-contact-at-a-time" change for simplicity, the human data revealed "composite transitions" where multiple contacts change simultaneously. These often appear in clusters, suggesting humans prioritize fluid motion over sequential stability.

Conclusion and Future Work

This research provides a quantified "timing budget" and a statistical roadmap for whole-body transitions. However, limitations remain. The current model identifies the presence of contact but does not yet account for the direction or magnitude of applied forces. Additionally, certain poses like elbow-supported leaning require more data to achieve statistical significance.

The ultimate objective is to develop a comprehensive "grammar of motion," enabling humanoid robots to navigate complex environments with the same grace and versatility as their human counterparts.

Summary updated

BibTeX
@inproceedings{mandery2015analyzing,
  author = {Christian Mandery and Júlia Borràs Sol and Mirjam Jöchner and Tamim Asfour},
  title = {{Analyzing whole-body pose transitions in multi-contact motions}},
  booktitle = {Humanoids},
  year = {2015},
  doi = {10.1109/humanoids.2015.7363479},
}
Figure