Welding robots stop splitting dissimilar-steel work by distance alone: thermal, stress and carbon coupling enter the objective, cutting simulated makespan and energy
Synopsis
Task allocation for collaborative robotic welding of dissimilar steel joints no longer divides work by robot-to-seam distance alone. The framework treats each robot as a player in an α-potential game and uses a physics-informed neural-network surrogate for repeated thermo-mechanical evaluation, so allocation accounts for residual stress and carbon-depleted-zone width. Across computational cases with 50 tasks and 2 robots, 100 tasks and 3 robots, and 200 tasks and 4 robots, it reduced normalized makespan and energy relative to distance-driven allocation while keeping lower predicted residual stress and carbon-depleted-zone width; these are computational predictions requiring independent experimental validation.
Interpretation
Task allocation for collaborative welding now writes physical coupling into the objective instead of minimizing robot travel alone. Distance-driven allocation does not explicitly account for neighboring welding operations changing the thermal state and mechanical response of later operations. Across the 50-task/2-robot, 100-task/3-robot and 200-task/4-robot cases, the framework reduced normalized makespan and energy relative to the distance-driven strategy while keeping lower predicted residual stress and carbon-depleted-zone width.
Each robot chooses tasks and order by its own utility, while the system measures overall cost with a common potential function. The α-potential game permits a bounded discrepancy between a single robot's utility change and the change in the common potential, with the bound summed from thermal, residual-stress and carbon-redistribution contributions. Under a finite feasible strategy space, a global maximizer of the potential is an α-Nash equilibrium, meaning no unilateral feasible deviation raises a robot's utility by more than α.
Repeated thermo-mechanical evaluations are carried by a physics-informed neural-network surrogate whose training loss includes transient heat conduction and mechanical equilibrium. A high-fidelity welding simulation costs far more than a geometric allocation calculation, making it expensive inside an optimization loop. The allocation procedure samples candidate deviations with log-linear learning, evaluates them with the surrogate, and periodically checks the selected allocation against the reference physical model.
Perspective
The conclusions rest on a computational study covering cases of 50 tasks with 2 robots, 100 tasks with 3 robots and 200 tasks with 4 robots, with heat input 0.8–1.2 kJ/mm, welding speed 5–10 mm/s, 3 passes, a maximum inter-pass temperature of 150 °C and shielding gas of 98% Ar plus 2% CO₂; predicted residual-stress and carbon-redistribution values need independent experimental validation before generalization to industrial welding.
α is not a universal material constant but an upper bound on the discrepancy introduced by the adopted interaction model, and must be estimated from the computational formulation; finite-time convergence results for log-linear learning hold only for exact potential games under their stated assumptions and do not transfer automatically to the α-potential case; carbon-depleted-zone predictions depend on the adopted diffusion and thermal-history model and should not be read as universal for all carbon-steel/stainless-steel combinations; surrogate generalization is limited by the domain covered in training simulations; the α-Nash result carries no universal convergence-rate or price-of-anarchy guarantee.