Grasp Planning from Human Prehension

Thea Iberall

Abstract

This paper presents a two-level hierarchical framework for robot grasp planning, drawing inspiration from human prehension. By mapping high-level task descriptions to "opposition spaces"—where "virtual fingers" apply force along specific axes—the model translates complex functional requirements into stable hand postures. This approach shifts the focus from purely geometric shape-matching to task-centric interaction, providing a robust architecture for dextrous robotic control.

Introduction

In traditional robotics, grasping is frequently modeled as a geometric problem: determining how a mechanical end-effector can physically encompass an object's shape. However, for multi-fingered, dextrous robot hands, the problem is significantly more complex. A single object can be held in an infinite number of ways, yet only a subset of these postures is appropriate for a given task. For instance, the grasp required to tap a nail with a hammer differs fundamentally from the grasp needed to swing it with force.

Thea Iberall's work addresses the critical gap between high-level strategic planning (e.g., "pick up the hammer to swing it") and low-level motor control. The goal is to develop a planner that simultaneously accounts for the physical constraints of the hand, the intrinsic properties of the object, and the functional requirements of the intended task.

Methodology: A Two-Level Hierarchical Planner

The proposed architecture decomposes the grasping problem into two primary phases, mimicking the cognitive and physical stages of human prehension.

The two-level hierarchical architecture for grasp planning, showing the flow from high-level commands to specific opposition space parameters.

Phase 1: Task Requirement Mapping

The system begins with a high-level command, such as GRASP(OBJECT, TASK_CONTEXT), and retrieves relevant data from an object database. It then generates "internal task requirements" by calculating the necessary forces and torques required to overcome the object's weight and the inertial forces associated with the intended movement.

Phase 2: Opposition Space Generation

These requirements are mapped into an Opposition Space. To manage the high dimensionality of a multi-fingered hand, the model employs the concept of "Virtual Fingers"—groupings of real fingers that act as a single functional unit. The planner identifies "opposition vectors," which are the axes along which these virtual fingers must apply force. Three basic types of opposition are defined:

  • Pad Opposition: Finger pads pressing against each other.
  • Palm Opposition: Fingers pressing against the palm.
  • Side Opposition: The thumb pressing against the side of a finger.

Key Insights: Task-Centric Interaction

A central tenet of this framework is that a grasp is not merely a property of an object's geometry, but a property of the interaction between the hand and the object. The planner must negotiate various constraints, including anatomical limits of the hand, object properties (weight, texture), and environmental accessibility.

Comparison of hand preshaping and grasping for two different tasks, illustrating how task context dictates the functional posture.

By grouping fingers into virtual units, the system simplifies the control problem. Instead of solving for dozens of degrees of freedom, the tactical controller focuses on maintaining the opposition vector. For example, to prevent unwanted rotation during a lift, the opposition vector must pass through or near the object's center of mass.

Case Study and Results

The framework was demonstrated using a symbolic representation of a steel hammer. By defining the hammer's dimensions, weight, and surface features, the planner could successfully derive the physical parameters (magnitude, location, and direction of force) required for different tasks.

The primary contribution of this work is conceptual:

  1. Decomposition Architecture: A clear path from abstract commands to physical execution parameters.
  2. Opposition Space Parameters: A standardized set of variables (e.g., magnitude, location-on-object) for tactical robot control.
  3. Constraint Taxonomy: A comprehensive identification of the factors—anatomical, interactional, and environmental—that a robust planner must address.

Discussion and Future Work

While the framework provides a robust theoretical foundation, several limitations remain. The current model considers the hand in isolation, without accounting for how arm and body posture might constrain reachable orientations. Furthermore, the focus is on planning rather than execution; real-time sensory feedback (e.g., tactile data to detect slipping) is not yet integrated. Future research aims to verify these constraints through human studies and integrate the planner into full-body robotic systems to achieve true human-like dexterity.

Summary updated

BibTeX
@inproceedings{iberall1987grasp,
  author = {Thea Iberall},
  title = {{Grasp Planning from Human Prehension}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1987},
}
Figure