Source-linked AI summary
A Modular Dual-Arm Robotic Cell for Disassembly and Repair of Industrial Control Electronics
Maximilian Ruhe, Fabian Harlacher, Christian Friedrich, Martin Kipfmueller
TL;DR
Industrial control electronics are routinely repaired but were not designed for automated disassembly, which must handle variable device condition and contact-rich tasks. The paper integrates device-graph planning, dual-arm skills, perception, and execution through a shared ROS 2 architecture. Integration feasibility is demonstrated, while quantitative evaluation remains ongoing.
Problem
Industrial control electronics have heterogeneous geometries, uncertain wear, incomplete documentation, and no robot-oriented service features, complicating repair-oriented disassembly.
Source: §1 · p. 1Method
Results
The demonstrated status establishes integration feasibility across the cell, while quantitative evaluation is ongoing.
Source: §8 · pp. 8–9Takeaways & Limitations
The architecture targets repair-oriented disassembly of legacy industrial devices by linking sequence knowledge, verification, and force-aware execution in one loop.
Source: §9 · p. 9Quantitative evaluation is ongoing, with planned measures spanning subtasks, complete cycles, perception, recovery, and sequence planning.
Source: §8 · p. 9
Abstract
from arXiv · showhide
Industrial control electronics such as programmable logic controllers, servo drives and operator panels are routinely repaired in plant maintenance, but were never designed for automated disassembly. This paper presents a modular dual-arm robotic cell for repair-oriented disassembly, using two collaborative manipulators, interchangeable tools, red-green-blue-depth (RGB-D) and wristlevel perception, force/torque sensing and a Robot Operating System (ROS) 2- based control with Behavior Tree (BT) execution, teleoperation, digital-twin support and bounded learning-based contact skills. The process is decomposed into sequence planning, symbolic execution with fallbacks, force-limited tool skills, visual condition assessment and demonstration-based adaptation. A CADderived device graph encodes the disassembly order, access constraints, tools, feasible removal directions and verification states and converts them into operation objects for the BT and motion layers. Grounded in three representative devices, the cell covers screw removal, damaged-fastener fallback, snap-fit opening, connector release, cooperative printed circuit board (PCB) extraction and condition-based repair decisions. The main contribution is an architecture linking sequence knowledge, perception, verification and force-aware skills through one ROS 2 interface across simulation, teleoperation and real hardware.
1 Introduction
Industrial control electronics are routinely repaired but were not designed for automated disassembly, which must handle uncertain device condition and contact-rich operations. The paper presents a repair-oriented dual-arm cell that combines structured task logic with adaptive behavior.
1 Introduction
- Repair disassembly must contend with heterogeneous geometries, uncertain wear, incomplete documentation, and devices altered by previous repairs.Operations include releasing snap-fits without breaking brittle plastic, removing PCBs without bending them, and unplugging ribbon connectors within force envelopes.Source: §1 · p. 1
- The cell combines device-family models, precedence relations, tool selection, safety envelopes, and Behavior Tree fallbacks with perception, teleoperation, and learned contact skills.The hardware, digital twin, and ROS 2 stack share a command and observation interface across simulation, teleoperation, and physical execution.Source: §1 · p. 2
- The paper presents a repair-oriented dual-arm cell and toolbox for screws, snap-fits, plug connections, PCBs, and inspection.Its contributions also include CAD-supported sequence planning, operation objects with target frames and verification criteria, and a shared ROS 2 interface.Source: §1 · p. 2
2 Related Work
Prior work addresses disassembly planning, execution, manipulation skills, and perception, but the paper identifies integration into a repair-oriented loop as the gap. It centers that integration on operation objects connecting sequence knowledge to executable behavior.
2 Related Work
- Reviews identify product variability, uncertain fasteners, contact-rich operations, and the laboratory-to-industry gap as open issues in robotic disassembly.Related efforts include battery disassembly and non-destructive disassembly of consumer and office electronics.Source: §2 · p. 2
- CAD-based relational models support disassembly sequence planning and removal directions, with later work addressing partial uncertainty and online replanning [10].Behavior Trees organize long-horizon execution and recovery, while SkiROS2 provides reusable symbolic capabilities.Source: §2 · p. 2
- Repair preservation is the organizing constraint: operations must be geometrically feasible, force-limited, and verifiable against device state.The operation object is positioned as the interface between sequence knowledge and executable robot behavior.Source: §2 · p. 3
3 Target Devices and Repair Requirements
Three industrial devices represent distinct disassembly challenges, from layered displays to brittle clips and screw-fastened electronics. Five design requirements specify repair-preserving access, flexible tools and resources, verification, and continuity across execution modes.
3 Target Devices and Repair Requirements
- The TP177B operator panel, S7-300 PLC, and Lenze EVS-9325 servo inverter represent layered HMI, compact controller, and servo-drive disassembly cases.Their structures include display modules and cable routing, brittle snap-fit housings, and screw-fastened covers with PCBs and edge-card connectors.Source: §3 · p. 3, Figure 1 · p. 3
- The five requirements are repair-preserving access, tool flexibility, bimanual resource reasoning, perception-driven verification, and data continuity across real execution, teleoperation, and simulation.They cover access to housings, boards, displays, and connectors, including damaged-fastener fallback and single-arm, fixture-assisted, or cooperative extraction.Source: §3 · pp. 3–4
4 Hardware Cell and Repair Toolbox
The cell uses two force-sensing manipulators for stabilization, tool work, and cooperative extraction. A reusable repair toolbox abstracts recurring operations into transferable skills parameterized for different devices.
4 Hardware Cell and Repair Toolbox
- Two 6-DOF UR15 manipulators share a workspace, with wrist-level force/torque sensing and 15 kg payload and 1300 mm reach per arm.One arm can stabilize a part while the other performs a tool operation, or both can act synchronously during extraction.Source: §4 · p. 4, Figure 2 · p. 4
- Bimanual handling provides a force-monitored alternative to dedicated clamping fixtures for fragile or variable parts.The text notes risks of cracking aged housings under over-constraint and bending large PCBs under asymmetric extraction forces.Source: §4 · p. 4
- The repair toolbox uses reusable operation primitives rather than device-specific procedures for fastening removal, connector handling, housing opening, extraction, and inspection.Skills are parameterized by CAD information, perception, and process context to support different industrial devices and geometries.Source: §4 · p. 4
5 Software Architecture
The ROS 2 architecture connects planning, execution, perception, teleoperation, learned skills and motion control through a shared command and observation path. A single controller manager routes commands to either the Isaac Sim twin or the physical cell.
5 Software Architecture
- The task layer combines sequence planning and Behavior Tree execution, then routes motion goals through MoveIt 2 and ros2_control to the twin or real cell.Perception, the CAD device graph, VR teleoperation and learned skills feed the task layer.Source: Figure 3 · p. 5, §5 · p. 5
- One controller manager switches between the Isaac Sim digital twin and the UR15 arms and Schunk grippers in the real cell.The twin connects through the ROS 2 bridge; the arms use RTDE and the grippers Modbus/TCP.Source: Figure 3 · p. 5
- The architecture shares command and observation messages across simulation and hardware, while force/torque and perception feed back to skills and verification.The feedback supports local plan, move, sense and verify loops; contact-sensitive parameters may still require recalibration on hardware.Source: §5 · pp. 5–6
6 CAD-Supported Sequence Planning and Operation Objects
CAD-derived device graphs encode removal constraints and dependencies, while runtime observations and execution outcomes update the plan. The planner converts reachable tasks into operation objects for Behavior Tree execution and replanning.
6 CAD-Supported Sequence Planning and Operation Objects
- The CAD-derived graph encodes component relations, removal spaces, dependencies, tool access, approach directions, force limits and verification requirements.Runtime scene information updates accessibility, component state and verification labels; deviations can trigger replanning.Source: §6 · p. 6
- Each operation object carries task-specific execution constraints, such as a snap-fit target, lever, insertion direction and admissible force range or PCB grasp regions and extraction axis.The graph also records relevant component-specific information, including screw pose and head type, clip lever direction, and PCB grasp regions.Source: §6 · pp. 6–7
- A scheduling tick selects reachable graph operations, attaches perception labels and confidence, allocates tools and resources, and dispatches them for execution.Returned outcomes update the graph and can trigger a fallback or new planning tick.Source: §6 · p. 7
7 Execution, Perception and Adaptive Skills
Behavior Tree subtrees execute operations with precondition checks, verification and fallbacks, while perception and adaptive skills address uncertain contact and condition assessment. Simulation supports development before hardware calibration.
7 Execution, Perception and Adaptive Skills
- Behavior Tree subtrees combine precondition checks, tool selection, motion planning, skill execution, postcondition verification and fallback handling.Bit slip or a damaged screw head can dispatch milling; excessive levering force stops the skill and requests re-approach.Source: §7 · p. 7
- Two-stage perception localizes devices and candidate parts with RGB-D, then uses wrist-level views to refine the active target and assess PCB defects.Defect findings update operation and verification states and can switch the process to targeted repair while leaving healthy modules assembled.Source: §7 · p. 7, Figure 4 · p. 8
- Isaac Sim supports reachability and collision checks, viewpoint planning, Behavior Tree debugging and synthetic data generation through the ROS 2 interface shared with the physical cell.Contact-rich skills are pre-trained under randomized conditions in the twin before hardware calibration.Source: §7 · p. 7
- Teleoperation captures demonstrations, and learned contact skills are used where geometry alone is insufficient; Adaptive Compliance Policy combines visual, proprioceptive and force/torque observations.The policy predicts reference and virtual target poses plus stiffness, allowing compliance to vary with contact needs.Source: §7 · p. 8
8 Integration, Demonstration and Evaluation Plan
The cell integrates planning, Behavior Tree execution and perception across three device stress cases, with demonstrations of several disassembly operations. The paper reports integration feasibility while quantitative evaluation remains ongoing.
8 Integration, Demonstration and Evaluation Plan
- The S7-300, Lenze servo drive and TP177B test snap-fit opening, cooperative PCB extraction and layered display disassembly, respectively.The graph selects reachable operations, the Behavior Tree checks postconditions, and perception informs whether to continue, inspect, recover or enter a repair branch.Source: §8 · p. 8
- Demonstrations include screw detection and removal, cover opening, learned-compliance snap-fit access, teleoperated intervention and demonstration capture.The reported status demonstrates integration feasibility; quantitative evaluation is ongoing.Source: §8 · pp. 8–9
- Planned evaluation covers subtask and complete-cycle performance, including success, perception, force asymmetry, tool-change and cycle-time metrics against scripted, teleoperated and learned-skill baselines.The stated limitations include recurring device-family dependence, initial graph creation and sim-to-real calibration for aged plastic, friction and damaged fasteners.Source: §8 · p. 9
9 Conclusion
The paper concludes with a modular dual-arm cell that integrates device-graph reasoning, force-aware execution, verification, teleoperation, and digital-twin development across simulation and physical hardware. It targets legacy industrial devices repaired in maintenance workflows and aims to support their repair within circular factory workflows.
9 Conclusion
- The cell links device-graph reasoning, force-aware Behavior Tree execution, perception-driven verification, teleoperation, and digital-twin development across simulation and physical hardware.The integration is organized around operation objects in a repair-oriented loop.Source: §9 · p. 9
- The architecture addresses legacy industrial devices that are routinely repaired in maintenance but were not engineered for automated disassembly.Source: §9 · p. 9
- The cell targets a high-value segment of the repair chain by preserving embedded material and manufacturing effort and offering a scalable pathway into circular factory workflows.Source: §9 · p. 9