FORGE: Bridging General Time-Series and Motion Foundation Models for Few-Shot Motion Time Series Classification

Few-shot motion time series classification must exploit very limited labels while reusing knowledge from large-scale pretraining. General-purpose time-series foundation models (TSFMs) provide broad temporal priors, whereas motion foundation models encode motion-specific semantics, but integrating these heterogeneous representations is non-trivial. We propose FORGE, a dual-foundation framework that keeps both pretrained encoders frozen, decomposes their representations into shared transferable components and foundation-specific residuals, and integrates them through global shared routing followed by local residual refinement. Across 18 real-world motion datasets, FORGE consistently improves few-shot classification and also strengthens complementary zero-shot cross-dataset transfer. At 20 shots, the best configuration reaches 82.54% Accuracy and 83.06% Macro-F1, outperforming the strongest standalone motion foundation by 16.36 and 17.46 points, respectively.