Tell us about your data requirements and we'll get back to you within 24 hours.
Unscripted human behavior captured through frontier multimodal sensors.
Real-world structure transferred into simulation. Scene geometry from observation, not generation.
Human data to robot policy to physical deployment. One partner, end to end.
Coactivation-Assembled Weight Layers
We propose Coactivation-Assembled Weight Layers (CAWL), a transformer-compatible linear block with effective map Weff(c) = U A(c) Vᵀ, where A(c) is a sparse latent adjacency assembled from coactivated latent neurons. The layer routes by familiarity: a similarity query over latent neuron embeddings computes a gate that interpolates between circuit recall from a persistent table and circuit construction by dynamic pair modules.
Paper not yet publicly available.
CoRL 2026
