Trajectory Representation Learning

It studies how to encode paths—such as vehicle movement, robot actions, and agent interaction traces—so useful structure survives compression and supports downstream tasks. Common approaches include self-supervised contrastive and masked-sequence pretraining, graph and hierarchical encoders, spatial tokenization, and continuous latent or generative models for waypoints, motion phases, and trajectory-level effects. These representations are tested on similarity search, classification, travel-time prediction, driving forecasts, and robot manipulation or navigation across tasks and embodiments. A central tension is balancing compact, reusable embeddings with enough temporal, geometric, and contextual detail for reliable long-horizon prediction and control in new scenes or at different execution rates.