Data-driven trajectory performance prediction for industrial robots via multi-modal retrieval

Abstract Evaluating industrial robot performance typically involves costly individual tests or standardized procedures that cover only a limited set of motion sequences. This makes it difficult to transfer measured accuracy to new trajectories, especially when robot configurations or workspace regions change. To enable data-driven performance prediction at scale, this work introduces a publicly available database of over 30,000 measured trajectories of an ABB IRB4400 industrial robot. Based on this database, a multi-modal retrieval framework is proposed that identifies similar motion sequences and predicts trajectory performance from their measured deviations. The framework employs a two-stage pipeline: embedding-based candidate retrieval across multiple motion modalities followed by constrained Dynamic Time Warping refinement for precise spatio-temporal alignment. Experimental evaluation on 500 randomly sampled trajectories demonstrates that multi-modal retrieval outperforms a scalar-feature nearest-neighbor baseline and a supervised neural-network baseline. Segment-level decomposition further improves both prediction accuracy and computational efficiency. The framework operates without parametric model fitting and can incorporate additional measurements without structural modification, allowing predictive capability to grow with increasing database coverage.