01Robot world models
RoboJEPA trains a latent world model across 12 robot embodiments and reports that imagination error follows a predictable curve as compute grows. Larger versions preserve objects and contact for longer, while the authors connect that error to downstream planning performance.
Why it mattersIf the scaling relationship survives broader testing, teams can estimate world-model progress before every expensive real-robot evaluation.
Keep in mind: Scaling fits can break outside the measured regime.
Read the source · Zholus, Artem and colleagues ↗023D robot action
UNITAS represents robot actions and the resulting scene changes together in metric 3D rather than as pixels. The 1.7-billion-parameter model reports strong manipulation comparisons and shows real-robot sequences, but not open-world autonomy.
Why it mattersA 3D-native model can reason about physical distances and scene motion directly instead of treating the world as view-dependent pixels.
Keep in mind: Leading results are relative to the selected comparisons.
Read the source · Wang, Ruixiang and colleagues ↗03Tactile robotics
TacHair watches a high-resolution tactile contact map while a robot strokes a hair-covered headform, adding small corrections when contact drifts. The authors report 525 real-robot trials on two platforms; people and autonomous grooming were not tested.
Why it mattersDeformable hair can stay in contact while drifting under the tool; explicit contact-distribution feedback addresses a failure that force or trajectory control can miss.
Keep in mind: Trials use secured headforms rather than people.
Read the source · Hu, Ruiyi and colleagues ↗04Robot simulation
USDCraft turns measured or generated object geometry into editable articulated OpenUSD programs. The paper shows drawers, toasters and other assets moving through real-to-sim-to-real manipulation, while geometric validation remains necessary.
Why it mattersFunctional joints and collision geometry are the difference between a pretty mesh and an asset useful for robot simulation.
Keep in mind: Real-to-sim-to-real results cover a limited object set.
Read the source · Zhang, Chuanrui and colleagues ↗05Cross-embodiment learning
LAC-WM learns one latent action space from human and robot datasets instead of a separate label system for every body. It reports better adaptation to unseen embodiments, including gains over explicit action conditioning in the tested manipulation settings.
Why it mattersA shared action representation could let robot world models reuse more data across bodies with incompatible control spaces.
Keep in mind: The headline gain is benchmark-specific.
Read the source · Huang, Huang and colleagues ↗06Interactive navigation
Path-Creative Navigation lets a robot decide when a door, box or chair should be moved rather than treated as permanent geometry. The strongest evidence is simulation, but it makes a useful distinction between actionable blockers and ordinary obstacles.
Why it mattersReal navigation sometimes requires changing the environment, not merely planning around it.
Keep in mind: Most evidence is simulated.
Read the source · Xi, Haoyu and colleagues ↗07Robot policy evaluation
A new analysis argues that simple robot-action regression often loses to flow or diffusion policies because it models residual error poorly, not only because demonstrations are multimodal. Residual corrections narrow much of that gap across simulations and three reported real-robot tasks.
Why it mattersTeams may recover strong manipulation performance without defaulting to heavier generative action heads where the residual assumptions hold.
Keep in mind: It does not dominate every flow or diffusion policy.
Read the source · Zhou, Yuchen and colleagues ↗