A whole-body pose taxonomy for loco-manipulation tasks

Júlia Borràs Sol, Tamim Asfour

Humanoid robots have achieved significant success in bipedal locomotion on flat, unobstructed surfaces. However, real-world environments—cluttered rooms, industrial sites, or stairs—require robots to interact with their surroundings to maintain stability. When a robot leans against a wall, crawls under an obstacle, or uses a handrail, traditional balance algorithms like the Zero Moment Point (ZMP) often become insufficient. A new research paper from the Karlsruhe Institute of Technology (KIT) proposes a "Whole-Body Pose Taxonomy" to standardize how robots use their entire bodies for support, treating balance as a form of "grasping the world."

The Challenge of Multi-Contact Stability

The core difficulty in "loco-manipulation"—tasks involving simultaneous locomotion and environmental interaction—is the lack of a unified vocabulary. While robotic hands benefit from well-established "grasp taxonomies" (e.g., power vs. precision grasps), whole-body movements have historically lacked such classification. Without a framework to categorize poses, planning complex motions with multiple contact points is computationally expensive and difficult to generalize. Currently, robots often struggle to autonomously decide whether to lean, step, or grab a rail to stay stable while performing a task.

Defining the Whole-Body Pose Taxonomy

The authors propose a taxonomy of 46 distinct whole-body poses, derived from three primary categories:

  1. Standing: Support primarily through feet and/or hands.
  2. Kneeling: Support involving one or both knees.
  3. Resting: High-stability poses where the torso or large body areas contact the environment, typically resulting in low mobility.

The taxonomy is organized by two primary criteria: the number of supports (ranging from single support to quadruple support) and the type of contact. The framework simplifies complex physical interactions into five fundamental contact types: hold (bilateral), palm, arm, feet, and knee.

The taxonomy classifies interactions based on specific contact types, ranging from precise finger-tip contact to broad knee or arm support.

Mapping Transitions and Motion Primitives

Beyond static poses, the framework defines how a robot can transition between states. By representing the taxonomy as a graph where edges connect poses reachable by changing exactly one contact point, the researchers create a roadmap for autonomous motion planning. This allows for the "semantic segmentation" of complex tasks into discrete motion primitives. This segmentation helps a robot distinguish between the part of an action meant for moving (locomotion) and the part meant for doing work (manipulation).

A sequence of poses (2.3, 3.4, 3.1, 2.1) illustrating how a robot transitions between different support configurations during a task.

Experimental Validation

To validate the taxonomy, the researchers analyzed human motion data from the KIT whole-body human motion database. They focused on a sequence of a human climbing stairs while utilizing a handrail. By applying a velocity threshold ($0.15 m/s$) and collision detection, the system successfully mapped the 5.5-second motion onto the transition graph.

The results demonstrated that while walking cycles are typically symmetric, the introduction of environmental supports like handrails creates an asymmetric "support graph." Identifying these patterns is crucial for robots to mimic human-like stability and efficiency in complex terrains. The system accurately identified exactly how the human moved from double-foot support to foot-plus-handrail support and back.

Discussion and Future Directions

The current taxonomy of 46 poses covers the most common robotic configurations, though the authors note that over 190 combinations are theoretically possible. Future work will involve expanding the taxonomy to include dynamic transitions—where multiple contacts change simultaneously, such as during running or jumping—and integrating the framework into autonomous decision-making systems. By treating the environment as an object to be "grasped" by the whole body, robots can move closer to the versatile mobility seen in humans, instinctively reaching for a wall or rail when stability is threatened.

Summary updated

BibTeX
@inproceedings{sol2015whole,
  author = {Júlia Borràs Sol and Tamim Asfour},
  title = {{A whole-body pose taxonomy for loco-manipulation tasks}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2015},
  doi = {10.1109/iros.2015.7353578},
}
Figure