A Composable Framework for Policy Design, Learning, and Transfer Toward Safe and Efficient Industrial Insertion

Rui Chen, Chenxi Wang, Tianhao Wei, Changliu Liu

Abstract

Delicate industrial insertion tasks, such as PC board assembly, remain challenging for industrial robots due to low error tolerance, component fragility, and significant task variations. To deliver a feasible robotic solution, systems must account for hardware limits and minimize integration effort. This paper proposes a composable framework for the efficient integration of a safe insertion policy on existing robotic platforms. The policy features an interpretable modularized design that can be learned efficiently on hardware and transferred to new tasks with ease. The framework includes a safe insertion agent (SIA) as a baseline policy, an optimal configurable Cartesian tracker (OCCT) for hardware interfacing, a probabilistic inference module (PIM) to handle component variety, and a safe learning module (SLM) to optimize parameters. Experimental results on a UR10 robot demonstrate that this framework achieves high safety, accuracy, and robustness while maintaining data and time efficiency during learning.

Introduction

In industrial automation, "insertion"—the process of mating components like IC chips with sockets—is a fundamental yet demanding task. When dealing with delicate components such as integrated circuit (IC) chips with numerous fragile pins, the margin for error is nearly zero. Traditional robotic approaches often struggle with four primary challenges: low error tolerance where sub-millimeter misalignments cause failure, perception noise from imperfect sensors, hardware latency in communication, and the difficulty of transferring policies across varying tasks.

To address these issues, we present a composable framework that decouples high-level logic from low-level control. This modularity allows for safer execution and more efficient learning, ensuring that robots can perform high-precision assembly without damaging expensive hardware or components.

The Composable Framework

The core of our approach is the division of the control policy into four distinct, specialized modules:

  • Safe Insertion Agent (SIA): This module acts as a logic-based state machine. It defines the high-level routine—alignment, insertion, and a "lift-and-retry" logic if a collision is detected. By providing an interpretable safety baseline, the SIA ensures the robot never applies excessive force.
  • Optimal Configurable Cartesian Tracker (OCCT): Serving as the interface to the robot hardware, the OCCT uses a Linear Quadratic Regulator (LQR) to plan smooth trajectories. It is specifically tuned to handle the communication delays inherent in ROS-based industrial systems, ensuring precise movement.
  • Probabilistic Inference Module (PIM): To handle uncertainty in component placement, the PIM utilizes a probabilistic model. If an insertion fails, the PIM analyzes the collision coordinates to infer a more accurate insertion pose for subsequent attempts.
  • Safe Learning Module (SLM): The SLM serves as the global optimizer, tuning parameters across the other modules using Covariance Matrix Adaptation Evolution Strategy (CMA-ES). This derivative-free optimization is ideal for the non-differentiable performance metrics found in physical contact tasks.

Methodology and Sim-to-Real Strategy

Our framework employs a strategic "sim-to-real" approach to minimize training time on physical hardware. The PIM is trained in a custom simulation environment where thousands of "defective" or misaligned chip poses can be generated safely. However, because contact physics are notoriously difficult to simulate accurately, the OCCT and SIA are tuned directly on the physical hardware. This ensures the low-level control is optimized for the actual dynamics of the robot, bypassing the "reality gap."

Furthermore, we utilize parameter decomposition. While parameters within a single module (like tracking speed and alignment offset) are often coupled, parameters between different modules remain relatively independent. This allows us to optimize groups of parameters separately, significantly reducing the dimensionality of the learning problem.

Experimental Evaluation

We evaluated the framework using a UR10 robotic arm equipped with an Ensenso X36 stereo camera and a force-torque sensor. The task involved inserting 2x4 pin IC sockets into printed circuit boards (PCBs).

Key Results

  • Task Efficiency: The PIM-guided system required an average of only 2.8 insertion attempts to succeed, even when starting from inaccurate initial poses. This significantly outperforms standard random search methods.
  • Safety and Precision: The OCCT maintained a horizontal tracking error of less than 0.1 mm. During collisions, the system limited instantaneous force to under 20 N, with forces exceeding 5 N lasting only for milliseconds—well within the safety threshold for fragile IC pins.
  • Learning Speed: The system is highly efficient to deploy. Learning for the OCCT and SIA took only 30 minutes on hardware. While the PIM simulation required 180 minutes, the final fine-tuning of the integrated system on real hardware was completed in just 60 minutes.

Conclusion

This composable framework provides a robust solution for delicate industrial assembly. By separating high-level probabilistic inference from low-level hardware tracking, we achieve a system that is both safe and easy to transfer between different robotic platforms. Future work will focus on expanding the SIA and PIM modules to accommodate a wider variety of assembly tasks beyond multi-pin insertions and conducting further theoretical analysis of the system's robustness under extreme environmental variations.

Summary updated

Abstract

Delicate industrial insertion tasks (e.g., PC board assembly) remain challenging for industrial robots. The challenges include low error tolerance, delicacy of the components, and large task variations with respect to the components to be inserted. To deliver a feasible robotic solution for these insertion tasks, we also need to account for hardware limits of existing robotic systems and minimize the integration effort. This paper proposes a composable framework for efficient integration of a safe insertion policy on existing robotic platforms to accomplish these insertion tasks. The policy has an interpretable modularized design and can be learned efficiently on hardware and transferred to new tasks easily. In particular, the policy includes a safe insertion agent as a baseline policy for insertion, an optimal configurable Cartesian tracker as an interface to robot hardware, a probabilistic inference module to handle component variety and insertion errors, and a safe learning module to optimize the parameters in the aforementioned modules to achieve the best performance on designated hardware. The experiment results on a UR10 robot show that the proposed framework achieves safety (for the delicacy of components), accuracy (for low tolerance), robustness (against perception error and component defection), adaptability and transferability (for task variations), as well as task efficiency during execution plus data and time efficiency during learning.
BibTeX
@inproceedings{chen2022composable,
  author = {Rui Chen and Chenxi Wang and Tianhao Wei and Changliu Liu},
  title = {{A Composable Framework for Policy Design, Learning, and Transfer Toward Safe and Efficient Industrial Insertion}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2022},
  url = {https://openreview.net/forum?id=U0Hv9RX14tl},
}
Figure