Collaborative research: anticipatory robotics: world models and generative planning for safe, human-aligned autonomy

The National Science Foundation Division of Computer and Network Systems awarded Case Western Reserve University $401,510 on October 1, 2026, under the Computer and Information Science and Engineering program (CFDA 47.070) to develop robotic welding systems capable of anticipatory autonomy and adaptive task performance.

The research combines sensing, prediction, and planning to enable welding robots to learn from expert demonstrations and adapt across variable joint configurations, materials, and fixtures—addressing the nation's shortage of skilled welders and the brittleness of today's pre-programmed systems. The project establishes embodied multimodal perception methods that fuse optical, infrared, electrical, acoustic, and inertial signals into a physics-grounded latent state representation, while constructing a differentiable 3D scene model to capture weld pool dynamics, arc evolution, and manipulation trajectories. The work serves as a testbed for anticipatory robot autonomy in complex, partially observed physical interaction. Expected benefits include stronger U.S. manufacturing resilience, higher production throughput, improved weld quality and consistency, reduced rework and material waste, safer operation in hazardous environments, and new educational tools for AI-enabled manufacturing training.

Work is performed at Case Western Reserve University, Office of Research Administration (operating in affiliation with University Hospitals of Cleveland) in Cleveland, Ohio. The period of performance runs from October 1, 2026, through September 30, 2029. This is a Project Grant assistance type.

Abstract

Welding holds together cars, ships, energy systems, and critical infrastructure, but the nation faces a shortage of skilled welders and many welding jobs remain hazardous, variable, and difficult to automate. Today?s welding robots are powerful but brittle: they usually repeat pre-programmed paths and fixed parameters, and they often fail when the joint, material, fixture, or desired welding style changes. This project asks a new scientific question: can a robot learn to anticipate how a weld will evolve before IT acts, much like an expert welder mentally predicts heat flow, ARC behavior, and bead formation? The research will create robotic welding systems that combine sensing, prediction, and planning so they can adapt across tasks, learn from expert demonstrations, and make safer decisions. The expected benefits include stronger and more resilient U.S. Manufacturing, higher production throughput, improved weld quality and consistency, reduced rework and material waste, safer operation in hazardous environments, and new educational tools for training students and workers in AI-enabled manufacturing. The proposed approach introduces an embodied, dynamic perception pipeline that fuses multi-rate sensor streams into a unified, temporally grounded, embodied latent representation, while simultaneously constructing a differentiable 3D scene model capable of capturing weld pool dynamics, ARC evolution, and manipulation trajectories. The project will establish a general framework for anticipatory robot autonomy using welding as a demanding testbed for complex, partially observed physical interaction. The research will first develop embodied multimodal perception methods that combine optical, infrared, electrical, acoustic, and inertial signals to infer a compact, physics-grounded latent state of the evolving weld scene. Dynamic three-dimentional gaussian representations will connect sensor fusion with geometric and thermal state estimation under ARC glare, plume occlusion, and rapidly changing material behavior. The project will then develop context-conditioned world models that learn how latent weld states evolve under different actions, machines, materials, and operator styles. These models will support multi-step predictive rollouts, quantify uncertainty, and embed physical constraints and human preference structure into the robot?s internal simulation. Building on this predictive capability, the research will create generative planning algorithms in which diffusion- or flow-based policies generate candidate action trajectories, evaluate them through world-model rollouts, and enforce differentiable safety constraints. The resulting system will be evaluated across multiple ARC welding processes on multiple robot platforms, while benchmark datasets and held-out transfer tests will measure generalization beyond the conditions used for training. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the foundation's intellectual merit and broader impacts review criteria.- subawards are not planned for this award.