The last populated daily window emphasized efficient use of scarce action signals. Today’s papers keep that concern and add a stronger signal around predictive supervision and explicit geometry.
Robot-learning teams can make predictive supervision more useful by expressing future changes in control-aligned coordinates, checking synthetic trajectories with explicit multi-view geometry, and using action-free…
Robot work dominates this day. Vision-language-action (VLA) papers focus on making policies cheaper, more geometry-aware, and safer to run on hardware.
Robot labs can improve deployment readiness by adding failure alarms to rollout harnesses, pruning VLA layers before downstream fine-tuning, and repairing object-specific failures with 3D-consistent augmented episodes.
This day is about world models becoming the control surface for robotics. DIAL ties a VLM to action through a latent future state and claims a large data-efficiency gain in VLA training.
A latent future representation is emerging as a concrete control interface for manipulation models, with reported gains in demonstration efficiency and early signs that mixed human-plus-robot data can support transfer…