Class-Incremental Motion Forecasting

A team of researchers has introduced a new framework for autonomous vehicle perception called class-incremental motion forecasting, designed to predict the trajectories of moving objects even as new, previously unseen categories emerge over time [1]. The approach, detailed in a paper revised on 18 June 2026, addresses a core limitation in current motion forecasting systems, which typically operate in a closed-world setting with a fixed set of object types and an assumption of flawless sensor data [1][2]. The proposed framework, named cOntinual Motion PrEdictioN (OMEN), is the first end-to-end system built to sequentially learn new object classes from camera images while retaining performance on previously learned ones, a challenge known as mitigating catastrophic forgetting [2][3]. Nicolas Schischka is listed as the corresponding author on the submission [1]. To adapt to new classes without requiring full dataset storage or re-training, the method generates motion forecasting pseudo-labels for known object categories. These are then matched with 2D instance masks produced by an open-vocabulary segmentation model [1][2]. A 3D-to-2D keypoint voting mechanism is employed to filter out inconsistent and overconfident predictions, improving reliability [1][3]. In parallel, a query feature variance-based replay strategy selects informative past sequences to help the model preserve prior knowledge [2][4]. The framework was evaluated on the nuScenes and Argoverse 2 datasets, where it successfully maintained performance on known classes while adapting to novel ones [1][3]. The researchers also demonstrated zero-shot transfer to real-world driving scenarios and extended the method to open- and closed-loop end-to-end class-incremental planning on nuScenes and NeuroNCAP [1][4]. The paper was first submitted on 10 March 2026, with subsequent updates in March and June of that year [1].