Abstract
This paper explores the introduction of "task frames"—coordinate systems attached directly to manipulated objects—as a means to bridge the gap between high-level symbolic planning and low-level motor control in robotics. By defining actions within the object's own frame of reference, complex trajectories are reduced to simple, invariant commands, facilitating more efficient computation and execution.
Introduction: The Representation Gap
In the early 1980s, robotics research was bifurcated into high-level AI (symbolic logic) and low-level control engineering (servomechanisms). Symbolic planners often ignored geometric constraints, while servomechanisms operated in a fixed "world frame." This made tasks like turning a crank mathematically arduous because the required forces and positions changed constantly relative to the world. Ballard (1984) proposed a Geometric Level of abstraction to sit between these two domains.
Methodology: The Task Frame Hierarchy
The central innovation is the Task Frame, a coordinate system intrinsically fixed to the object being manipulated. Ballard defines a three-level hierarchy:
- Symbolic Level: Handles logic and high-level planning.
- Geometric Level: Manages task frames and spatial representations.
- Servomechanism Level: Executes inverse dynamics and joint torques.
By specifying actions in the object's frame, complex trajectories become invariant commands. Ballard proposes a STRIPS-like syntax for geometric actions, defining them with parameters, preconditions, and stopping conditions. The theory is grounded in quaternion-based mathematics to transform between world ($x_w$) and task ($x_t$) frames. Additionally, the paper introduces proprioceptive self-calibration, where a robot "discovers" the task frame by sensing motion constraints through its own joints.
Dynamics and Implementation
To bridge the gap to hardware, force and torque vectors in the task frame are mapped directly to joint torques using inverse dynamics equations. This is typically demonstrated using a conceptual two-link planar manipulator model.
Key Insights and Results
The primary advantage of task frames is invariance. Decoupling force constraints from geometric ones allows a single geometric description (e.g., DOORCLOSING) to handle various door types and orientations. Furthermore, the framework integrates control loops by mapping high-level "until" conditions (like ARRESTED) to low-level sensor thresholds (like force spikes).
Limitations and Future Work
The framework is currently optimized for single-object manipulation and does not address global path planning or complex failure recovery. Representing environments with multiple interacting objects and incorporating detailed friction models remain challenges for future development.
Summary updated