Abstract
Robotic assembly of tight-fitting parts is often hindered by real-world uncertainties in sensing, control, and part manufacturing. This paper introduces AMP-CAD, a planning system that automatically generates assembly plans by accounting for these "slops." Unlike traditional configuration space methods, AMP-CAD utilizes a potential-field approach to identify collision-prone segments of a path and switches to force-torque guided "fine motions" when necessary. Experimental validation on industrial robots shows the system can successfully automate complex 6-DOF assembly tasks with an 85% success rate.
Introduction
In ideal simulations, tasks like inserting a peg into a hole are trivial. However, physical implementation faces three primary types of uncertainty that make "mating" tasks difficult: initial pose uncertainty (sensing errors), control uncertainty (mechanical drift), and model uncertainty (manufacturing tolerances).
When these combined uncertainties exceed the clearance between parts (e.g., a 0.01-inch gap), standard path planning fails. AMP-CAD addresses this by discovering which path segments are "collision-prone" and assigning appropriate sensory-feedback strategies to navigate them safely.
Methodology
The AMP-CAD system employs a two-phase planning architecture to bridge the gap between high-level CAD descriptions and low-level robot execution.
Phase 1: Geometric Planning
The system first ignores uncertainty to find a nominal path using a potential field representation. In this model, obstacles exert a "repulsive force" on the part, guiding the planner through the "valleys" of the potential field to find a collision-free path in a perfect environment.
Phase 2: Uncertainty Analysis
The system then "thickens" the parts to account for real-world deviations:
- Model Uncertainty: Handled via the Most Material Condition (MMC), assuming parts are at their maximum size and holes at their minimum.
- Pose/Control Uncertainty: Calculated as a "total deviation vector" that accumulates as the robot moves.
Collision testing is performed by integrating the potential field over the boundary of these thickened parts. If the potential exceeds a specific threshold, the segment is flagged as collision-prone. The planner then segments the path and selects a "fine motion" strategy from a library—such as torqvecsearch for torque-guided insertion—to handle the contact.
Key Insights
- Dynamic Uncertainty Modeling: Uncertainty is not treated as a static bubble but as an accumulating error. This allows the robot to move quickly and "bravely" at the start of a path, only slowing down for cautious force-feedback maneuvers when the error becomes critical.
- Contact as Information: Successfully mating parts using force feedback effectively "resets" the relative uncertainty between them to the physical clearance of the parts. This updated model allows the planner to proceed with renewed precision for subsequent steps.
- Efficient 6-DOF Reasoning: By using potential fields instead of Configuration Space (C-space), the system avoids the "dimensionality explosion" typically associated with modeling 6-DOF uncertainty and tolerances, making it computationally feasible for complex 3D rotations.
Experimental Results
The system was evaluated using custom CAD models of complex assemblies, such as a housing requiring the simultaneous meshing and mounting of gears. Physical verification was conducted on industrial robots, including the Cincinnati Milacron T3-726 and Puma 762.
- Success Rate: AMP-CAD produced feasible plans in approximately 85% of real-world trials.
- Automation: The system successfully transitioned from CAD descriptions to executable robot code (Common Lisp and C) without human intervention in strategy selection.
Limitations and Future Work
Despite its success, the system is limited by its Strategy Library; it cannot complete tasks requiring maneuvers not pre-defined in its dozen-strategy set. Additionally, a 15% failure rate persists due to sensor noise and "micro-collisions" during force control, which can lead to infinite error-recovery loops. Future work involves refining the calculation of "swept volumes" for rotating parts, which currently relies on sampling due to the complexity of transcendental functions.
Summary updated