Abstract
This paper presents a hybrid robotic architecture that integrates global geometric path planning with local, specialized controllers. By leveraging the local geometric context within a configuration space (C-space), the system effectively navigates the challenges of uncertain dynamics and intermittent contact inherent in non-prehensile manipulation tasks, such as pushing objects through complex environments.
Introduction
Robotic manipulation often involves "non-prehensile" tasks—actions like pushing where the robot does not maintain a firm grip on the object. These tasks are notoriously difficult due to uncertain dynamics (e.g., unknown friction distributions) and intermittent contact, where hitting an obstacle causes drastic, non-linear changes in object behavior.
Traditional path planners are often too brittle, relying on perfect geometric models, while reactive systems lack the global perspective needed to reach a goal without getting stuck. This research introduces a hybrid architecture that bridges this gap by using the "local context" of the environment to select appropriate control strategies.
Methods
The proposed system utilizes a two-layer architecture:
- Global Path Planning: The system constructs a 3D configuration space (C-space) representing the possible positions and orientations $(x, y, \theta)$ of the object. A collision-free path is then planned from the start to the goal.
- Local Control via Context: The core innovation lies in using the "local context"—the occupancy of neighboring cells in the tessellated C-space—as a sensor input. Instead of a single universal controller, the system maintains a library of specialized controllers. When the robot detects a specific local geometry (the context), it activates the corresponding controller, often learned from physics simulations.
Results and Discussion
The framework was evaluated using 30 different part shapes across 20 distinct manipulation tasks, including complex peg-in-hole assemblies performed entirely through pushing. The hybrid approach demonstrated significant robustness, successfully completing tasks despite injected sensing noise (up to 10 degrees of angular error). Unlike baseline methods, this system avoided "limit cycles" and "chatter," common failure modes in reactive robotics.
While the approach is highly effective for 2D and 3D tasks, scaling to robots with more degrees of freedom presents a "curse of dimensionality" challenge. Future work aims to address "trap states" and explore online adaptation, allowing controllers to refine their parameters during execution.
Summary updated