Introduction
In the context of Life Cycle Engineering, remanufacturing—the process of restoring used products to a like-new condition—is a critical pillar of sustainability. However, the initial stage of disassembly remains a significant bottleneck. Automating the disassembly of complex components like starter motors is fraught with challenges. Used products often arrive in unpredictable states of wear, and industrial clamping systems frequently occlude critical parts. Furthermore, standard 3D sensors typically provide only single-view point clouds, leaving much of the object's geometry invisible. Small, feature-scarce components like bolts are particularly difficult for traditional computer vision algorithms to locate accurately.
To address these issues, this research proposes a multi-stage pipeline that integrates deep learning with "prior knowledge" (high-quality reference models) to achieve precise 6D pose estimation for autonomous robot disassembly.
Methodology: The Multi-Stage Pipeline
The core of the proposed system is a five-stage pipeline that bridges the gap between partial real-world scans and complete geometric models. This process utilizes Point Cloud Transformers (PCT) for segmentation and classical geometric registration for alignment.
- Binary Segmentation: A PCT model is employed to isolate the motor from the background clamping system within the real-world scan.
- Part Segmentation: A second PCT model segments a "full model" (a high-quality scan or CAD model) into six semantic parts, including the housing and individual bolts.
- Registration: The system aligns the known full model to the partial real-world scan. This involves an initial rough alignment using Fast Point Feature Histograms (FPFH), followed by a refined three-step Iterative Closest Point (ICP) process.
- Bolt Clustering: Using the DBSCAN algorithm, the system groups points belonging to individual bolts to determine their 3D centers.
- Normal Alignment: By analyzing the geometry of the motor's main housing, the system calculates the motor's axis, which defines the 3D orientation of the bolts.
Data Generation and Transfer Learning
Industrial applications often suffer from a lack of labeled data. This study utilized only 18 physical starter motors. To overcome this, the researchers developed MotorFactory, a Blender-based tool to generate large synthetic datasets with automatic labeling. The models were pre-trained on this synthetic data and subsequently fine-tuned on the limited real-world dataset, demonstrating the efficacy of transfer learning in specialized industrial domains.
Key Insights: Prior Knowledge as a "Cheat Code"
The defining strategy of this approach is the use of prior knowledge. Instead of attempting to identify small, noisy bolts in a partial scan, the system identifies the entire motor. Once the motor's global pose is established, the precise positions and orientations of the bolts are "inherited" from the high-quality reference model. This allows the system to estimate the 6D pose of bolts even when they are completely occluded from the camera's view.
Furthermore, the Point Cloud Transformer (PCT) architecture proved superior to older models like PointNet++ in identifying small semantic features. In the registration phase, a tiered ICP approach—gradually decreasing the maximum matching distance—was essential for achieving sub-millimeter precision.
Results and Performance
The system demonstrated high accuracy and precision suitable for industrial workflows:
- Segmentation: The PCT model achieved an overall accuracy of 90.47% and a Mean Intersection over Union (mIoU) of 84.58%.
- Precision: The registration process reached a Root Mean Square Error (RMSE) of 0.076 mm.
- Efficiency: The optimized point-to-plane registration pipeline completed in approximately 6.39 seconds, balancing the need for high precision with the practical time constraints of disassembly tasks.
Conclusion and Future Work
This research demonstrates that integrating prior geometric knowledge with deep learning significantly improves 6D pose estimation in occluded, real-world environments. Future work will focus on expanding the dataset to include a wider variety of end-of-life conditions, such as rust and physical damage, and further optimizing the pipeline for dynamic, real-time robotic interactions.
Summary updated