TL;DR
This paper introduces PolyFit, a novel supervised learning framework for robotic peg-in-hole assembly that leverages Force/Torque (F/T) sensor data for precise extrinsic pose estimation and misalignment correction. By employing a sim-to-real adaptation strategy, PolyFit achieves robust performance on unseen polygon shapes, significantly improving success rates in real-world scenarios.
What Problem Does This Paper Address?
Robotic peg-in-hole assembly is a fundamental yet notoriously difficult task. Misalignments, often stemming from sensor inaccuracies or mechanical errors, can lead to insertion failures, jamming, and even damage to the robot or components. Traditional methods often rely on complex compliance control or heuristic algorithms that require manual calibration for specific part geometries and struggle to generalize to new shapes. While learning-based approaches have shown promise, vision-based methods are susceptible to environmental factors like lighting and occlusion, and existing force-based methods can be data-hungry and risk-prone when collecting real-world data. This paper aims to bridge this gap by developing a robust, data-efficient, and generalizable peg-in-hole assembly framework using F/T sensing.
Methods
PolyFit introduces a paradigm shift from reinforcement learning to a supervised learning approach for peg-in-hole assembly. The core of the framework is an F/T-based supervised learning model designed for 5-Degrees-of-Freedom (DoF) assembly.
- Extrinsic Pose Estimation: The system uses F/T data to estimate the precise extrinsic pose (position and orientation) of the peg relative to the hole. To overcome the ambiguity of F/T readings (where different poses can yield similar readings), a "multi-point contact" strategy is employed. This involves a sequence of controlled rotations to gather richer contact information.
- Sim-to-Real Adaptation: To address the challenges of real-world data acquisition, PolyFit utilizes a sim-to-real adaptation method. This involves creating a "sim-real paired dataset" where identical contact F/T and pose information is captured in both simulated and real environments. This paired data is used to align the models, enabling effective transfer of learned behaviors from simulation to the real world. Two adaptation techniques are explored: Data-Level Adaptation (DLA) and Feature-Level Adaptation (FLA).
- Network Architecture: A lightweight, multi-layer perceptron (MLP) based network is used for extrinsic pose estimation. It processes concatenated force and torque features from initial and multi-point contacts.
Datasets
- Simulated Dataset: A large dataset was generated in a simulated environment (Isaac Gym) comprising diverse peg-hole shapes (4-10 vertices) and extrinsic poses. This dataset includes corresponding contact F/T readings.
- Sim-Real Paired Dataset: For sim-to-real adaptation, a minimal set of sim-real paired data was collected. This dataset captures identical contact F/T and pose information from both simulated and real-world peg-in-hole interactions for polygons with 4-6 vertices.
Key Insights and Ideas
- F/T as a Primary Sensor: The paper demonstrates that F/T sensors, when used effectively with multi-point contact, can provide sufficient information for accurate extrinsic pose estimation in peg-in-hole tasks, mitigating the reliance on vision sensors and their associated environmental sensitivities.
- Multi-point Contact for Ambiguity Resolution: The insight that a single F/T reading can correspond to multiple poses is crucial. The multi-point contact strategy effectively resolves this ambiguity by gathering a richer set of contact data, leading to more reliable pose estimation.
- Data-Efficient Sim-to-Real: The proposed sim-to-real adaptation method, using a minimal sim-real paired dataset, is a significant contribution. It allows for effective transfer of learned policies to the real world without requiring extensive and costly real-world data collection.
- Generalization to Unseen Shapes: The framework's ability to generalize to polygon shapes not encountered during training is a key indicator of its robustness and practical applicability.
Key Results
- Simulation Performance: PolyFit achieved impressive success rates in simulation: 97.3% for seen polygon shapes and 96.3% for unseen shapes.
- Real-World Performance: In real-world evaluations, PolyFit demonstrated substantial success rates of 86.7% for seen shapes and 85.0% for unseen shapes.
- Comparison to Other Methods: The sim-to-real adapted PolyFit significantly outperformed methods without adaptation (38.3% for seen, 28.8% for unseen) and a standard spiral-search method (18.3% for seen, 17.5% for unseen) in real-world peg-in-hole tasks.
- Pose Estimation Accuracy: In real-world testing, the proposed method achieved significantly lower pose estimation errors compared to methods without adaptation, with the "Ours" method achieving the best results (POS: 0.48mm, ORI: 0.17° for seen; POS: 3.91mm, ORI: 1.85° for unseen).
Limitations and Future Work
- Limitations: While the paper highlights impressive results, the real-world success rates, though high, are not 100%. The MAE for pose estimation in real-world scenarios, even with adaptation, can still exceed the 1mm peg-hole tolerance in some cases, though successful assembly is still achieved within a few trials.
- Future Work: The authors suggest incorporating a closed-loop framework to further enhance robustness. They also propose extending the application of PolyFit to more complex assemblies, such as connectors and cables, indicating a clear path for future research and development.
Summary updated