One Small AI Model Targets Both Prediction and Path Planning

According to AI TIMES, a research team from DGIST and KAIST has developed a training technique that reduces performance loss while handling both surrounding motion prediction in crowded spaces and robot path planning with a single small AI model. The research results will be presented at ECCV 2026, which runs from September 10 to 12, and the team said it kept the model size unchanged while boosting performance on both tasks by applying separated parameter learning to reduce interference between functions. Until now, putting separate models in place for each function in real robots has been a heavy burden because of computing and memory constraints. In particular, the team viewed the problem of performance degradation caused by conflicting learning over internal resources when multiple functions are packed into one model as skill conflict and proposed a solution. In validation using representative data and additional experiments, the team confirmed a trend of improved motion prediction accuracy and path stability. The team said this approach could expand beyond robots handling crowded environments to physical AI more broadly as it moves through real-world spaces.