⚡ electrical engineering

Physics-Coupled Task Allocation for Collaborative Robotic Welding of Dissimilar Steel Joints Using an α-Potential Game and a Physics-Informed Neural Network Surrogate

This paper proposes a physics-coupled task allocation framework for collaborative robotic welding of dissimilar steel joints that integrates an $\alpha$ -potential game for distributed decision-making with a physics-informed neural network surrogate to optimize thermal interactions, energy consumption, and production time while accounting for residual stress and carbon redistribution.

Original authors: mohammad reza omidi

Published 2026-10-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: mohammad reza omidi

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of heavy industry, joining two different types of steel is a delicate operation. When engineers need to connect a strong, rust-resistant stainless steel to a standard carbon steel, they face a unique challenge. These materials do not just sit side by side; they react to heat in different ways. As the welding torch moves along the joint, the intense heat causes the metals to expand and contract at different rates. This mismatch creates invisible internal forces, known as residual stress, which can weaken the final product or cause it to warp. Furthermore, the heat can cause carbon atoms to migrate from one metal to the other, creating weak spots near the seam. To manage these risks, factories often use teams of robotic arms working together. However, simply telling robots to work as fast as possible or to stay far apart is not enough. If one robot welds a section too quickly, the heat it leaves behind can ruin the quality of the weld the next robot attempts to make nearby. The decision of which robot does what, and in what order, must account for these hidden physical interactions.

A researcher at the University of Bonab has developed a new way to solve this puzzle, treating the group of robots not just as machines, but as a team of players making decisions that affect one another. Instead of relying on simple rules like "keep the robots far apart," the new system uses a mathematical approach called a potential game. In this framework, every robot has its own goal, such as finishing its assigned tasks quickly, but it also contributes to a shared goal of making the strongest, most efficient weld possible. The system acknowledges that a robot's decision to move or weld changes the environment for the others, creating a ripple effect of heat and stress. To handle this complexity without getting bogged down by slow computer calculations, the researcher trained a special type of artificial intelligence. This AI acts as a fast, simplified model of the physical world, predicting how the steel will react to heat and stress in real-time. By combining this smart prediction tool with the game-theory approach, the system can find a schedule where the robots work together to minimize energy use, finish the job faster, and keep the internal stress of the metal as low as possible.

The study focused on a scenario where multiple robots were tasked with welding joints between dissimilar steels. The researchers compared their new method against older strategies that simply tried to minimize the distance robots traveled or used standard optimization techniques. In computer simulations involving anywhere from fifty to two hundred welding tasks, the new system consistently outperformed the traditional methods. It managed to reduce the total time the robots spent working and lowered the amount of energy consumed. More importantly, it produced welds with significantly lower predicted internal stress and less of the dangerous carbon migration that weakens dissimilar joints. The system achieved this by allowing the robots to "negotiate" their tasks. When one robot considered taking on a new job, the system checked not just how long that job would take, but how the heat from that job would affect the metal the other robots were about to weld. If a task would create too much heat for a neighbor, the system would reassign it to balance the load.

To ensure the results were not just lucky guesses, the researcher broke the system down to see which parts were doing the heavy lifting. They tested the system with and without the artificial intelligence model and with and without the game-theory approach. The results showed that both components were essential. The artificial intelligence provided the speed needed to run thousands of simulations quickly, while the game-theory framework ensured that the robots made decisions that were good for the whole group, not just for themselves. The study also tested how the system handled different weights of importance, such as prioritizing speed over energy or stress. The system remained robust across these changes, consistently finding a balanced solution. However, the researcher is careful to note that these results are currently limited to computer simulations. The numbers regarding stress levels and carbon movement are predictions generated by the model, not measurements taken from physical steel beams. Before this method can be used in a real factory, the predictions must be verified by welding actual joints and measuring the results with physical instruments.

The significance of this work lies in its shift from viewing robotic welding as a purely geometric problem to treating it as a physical one. Traditional methods often ignore the fact that heat travels and stress builds up, assuming that if robots stay out of each other's way, everything will be fine. This new approach recognizes that in the world of dissimilar metals, the history of the weld matters. The heat from a weld made minutes ago changes the material for the weld made next. By using a system that understands these physical connections, the researcher has shown that it is possible to coordinate multiple robots to produce higher quality results. The method does not require the robots to be smarter or faster; it simply requires them to be aware of the invisible forces they are creating. While the study confirms that this approach works in a digital environment, the path to the factory floor remains open. The next step is to take these computational predictions and test them against the messy, unpredictable reality of actual steel and fire, ensuring that the digital promise of a stronger, faster weld holds true in the physical world.

Technical Summary: Physics-Coupled Task Allocation for Collaborative Robotic Welding of Dissimilar Steel Joints

Problem Statement
Collaborative robotic welding of dissimilar steel joints (e.g., ferritic/low-alloy to austenitic stainless steels) presents a complex decision-making challenge. Unlike conventional geometric task allocation, which primarily minimizes travel distance and time, welding operations in dissimilar joints are physically coupled. Neighboring welding operations alter the thermal state, mechanical response, and metallurgical structure of subsequent operations. Specifically, non-uniform thermal expansion leads to significant residual stresses, while high-temperature exposure causes carbon migration, creating carbon-depleted and carbon-enriched zones that compromise mechanical properties. Existing allocation methods often fail to explicitly account for these thermo-mechanical and metallurgical interactions, leading to suboptimal production outcomes regarding quality, energy consumption, and distortion.

Methodology
The study proposes a unified framework that integrates game-theoretic task allocation with physics-informed machine learning. The methodology consists of three core components:

$\alpha$ -Potential Game Formulation:
The multi-robot system is modeled as an $\alpha$ -potential game. In this framework, collaborative robots are treated as players with individual strategies (task assignment and execution order). The global objective function ( $J$ ) minimizes a weighted sum of makespan, welding energy, maximum residual stress, and carbon-depleted zone width.
- Unlike exact potential games where individual utility changes perfectly mirror global potential changes, this formulation uses an $\alpha$ -potential to account for the bounded discrepancy ( $\alpha$ ) caused by non-additive thermal and metallurgical interactions.
- Theoretical analysis establishes that a global maximizer of the potential function constitutes an $\alpha$ -Nash equilibrium within a finite feasible strategy space.
- Log-linear learning is employed as a stochastic strategy-update mechanism to explore the solution space, allowing robots to sample candidate deviations based on their utility improvements.
Physics-Informed Neural Network (PINN) Surrogate:
To address the high computational cost of repeated thermo-mechanical evaluations required by the optimization loop, a PINN surrogate is developed.
- The PINN approximates the transient temperature field and mechanical equilibrium by incorporating the governing partial differential equations (heat conduction and thermo-elastic constitutive relations) directly into its loss function.
- This approach provides a reduced-cost approximation of the physical response, enabling rapid evaluation of candidate task allocations without relying solely on expensive high-fidelity simulations for every iteration.
Computational Framework:
The algorithm generates an initial feasible allocation, iteratively selects robots to explore unilateral deviations, evaluates candidates using the PINN, and updates the allocation based on log-linear probabilities. Periodic validation is performed against a reference physical model. The study evaluates scalability across three task scales (50, 100, and 200 tasks) with varying robot counts (2, 3, and 4).

Key Contributions
The paper identifies five principal contributions:

Formulation: The task allocation for multi-robot welding is formulated as an $\alpha$ -potential game, explicitly modeling the coupling between local robot decisions and global system objectives.
Interaction Modeling: Thermal, mechanical, and metallurgical interactions (residual stress and carbon redistribution) are represented through a bounded $\alpha$ term, acknowledging the non-linear and history-dependent nature of welding physics.
Theoretical Result: A proof is provided establishing that a global maximizer of the $\alpha$ -potential function is an $\alpha$ -Nash equilibrium under finite strategy spaces.
Surrogate Development: A PINN is developed to serve as a computationally efficient surrogate for repeated thermo-mechanical evaluations, integrating physical laws into the learning objective.
Comparative Analysis: The framework is computationally compared against distance-driven genetic allocation, genetic allocation with iterative local search, and Gaussian-process-assisted multi-agent search.

Results and Performance
In the computational study, the proposed framework demonstrated improvements over distance-driven baselines:

Efficiency: The method reduced normalized makespan and welding energy consumption.
Quality: The framework simultaneously maintained lower predicted maximum residual stress and reduced carbon-depleted zone widths.
Mechanism of Improvement: These gains are attributed to the game formulation allowing robots to account for the systemic impact of their decisions and the PINN reducing the computational burden of iterative evaluations.
Ablation: The study distinguishes the contributions of the $\alpha$ -potential correction and the PINN acceleration, showing that both components are necessary for the full performance of the framework.

Significance and Limitations
The paper positions this work as a step toward integrating process-level physics with distributed multi-robot decision-making. By treating welding tasks as physically coupled rather than purely geometric, the framework aims to improve the quality of multi-robot welding decisions.

However, the author maintains a modest stance regarding the practical application of these results:

Computational Nature: The reported numerical values are computational predictions. The paper explicitly states that these results must be independently validated experimentally before being generalized to industrial applications.
Model Dependence: The validity of the framework depends on the accuracy of the underlying physical model and the PINN surrogate. The $\alpha$ parameter is treated as a modeling bound estimated from the formulation, not a universal physical constant.
Scope: The study focuses on the algorithmic and theoretical integration of game theory and physics-informed learning, distinguishing theoretical equilibrium properties from numerical performance observations.

In conclusion, the study provides a theoretical and computational basis for physics-coupled task allocation, offering a method to balance production efficiency with the complex physical constraints of dissimilar steel welding, while emphasizing the need for future experimental verification.

Technical Summary: Physics-Coupled Task Allocation for Collaborative Robotic Welding of Dissimilar Steel Joints

More like this