Archetypal Sparse Transcoders Find Motion Features and Physical-Hazard Detectors in V-JEPA 2

As embodied systems are deployed in increasingly complex environments, it becomes necessary to monitor their physical safety from the inside. Sparse autoencoder (SAE) feature patterns are known to signal out-of-distribution operation and impending hallucination in vision and language transformers; we ask whether analogous probes can be built for the self-supervised video encoders now entering the physical-AI stack. We introduce an archetypal sparse transcoder for V-JEPA 2: a sparse decomposition that combines the layer-to-layer prediction objective of transcoders with the convex-hull decoder constraint of archetypal SAEs. Trained on the block-10 FFN of a frozen V-JEPA 2.1 ViT-B/16 using Something-Something v2, the transcoder reaches held-out explained variance 0.479 at L0 = 59, against a random-initialization control at −1.4×104. A label-distribution protocol over the 174 SSv2 motion templates surfaces features that track a common motion primitive across different objects, and a physical-hazard probe finds a statistically enriched sub-population that fires selectively on falls, drops, and collisions, 5.7× rarer in the untrained control. Over eight million held-out tokens per feature, these labels hold far beyond the top-activating clips (7.9× to 64× per-token enrichment), and preliminary multi-source training raises the dictionary’s alive fraction from 4.5% to 60%.