A Robot Vocabulary Maps the Work of Construction Crews
A preprint lays out 41 reusable activity primitives from occupational records and videos, then scripts selected ones on a robot arm and hand.
241–260
A preprint lays out 41 reusable activity primitives from occupational records and videos, then scripts selected ones on a robot arm and hand.
Preprint: Models trained with video-derived 3D LiDAR labels scored higher than models trained with SAM2 labels, but below full ground truth.
The exact result includes four Galois points and 24 non-Galois points, alongside tests for identifying related points on smooth quartic surfaces.
Preprint: A system uses one 3D-track representation for camera and object motion, while a distilled version reaches 20.6 frames per second in tests.
Preprint: A hydrogenic model predicts strong same-parity suppression, while mass asymmetry and spatial separation change the calculated coupling.
This preprint uses preliminary ALICE data to predict that the strongest production shifts from the oxygen side to the proton side in more peripheral collisions.
Preprint: A cached event representation outperformed named baselines on offline ranking and retrieval metrics in industrial logs.
A preprint reports an agent that chooses tools for each question and revisits weak evidence before answering.
A graph-based framework was tested on three trained networks; accuracy was nearly unchanged at small error settings, then fell steeply as the allowance grew.
Preprint tests episode reconstruction on 3,000 counterfactually constructed questions and reports higher accuracy with much smaller prompts.
Preprint: An AC/DC solver matched key wave and idealized convection tests, while its projected runtime was about 1.36 times the hydrostatic step.
Preprint: Conditional lower bounds are obtained for both versions of the model, but the full model's direct Agmon estimate loses uniform control as the time horizon grows.
Preprint finds that near-minimum discrepancy is tied to more uniform coverage under stated mathematical conditions.
Preprint: A mathematical analysis finds the classical data-processing rule can fail when a learning system is limited to a constrained model class.
Preprint: A proof covers sufficiently large trees and clique sizes at least logarithmic in tree order, while triangles and other small cases remain open.
An arXiv Preprint reports higher success for LM-X after training on heterogeneous robot trajectories, including failed policy rollouts, and testing it in simulation and on physical tasks.
Preprint reports an operator-certified whole-space interval for H₂⁺, while a three-centre test shows transfer of the method but not a certified whole-space lower bound.
Preprint modeling finds direct pointing emerges under intermediate foraging difficulty and can shift to a gravity-referenced code in many tested settings, but it does not reconstruct real bee evolution.
A preprint tests TacForcing on six simulated and six real-world tasks, with higher reported average success than comparison systems.
Preprint: A theoretical model predicts a momentum-space state that is neither fully localized nor evenly spread, but stronger-disorder behavior remains unresolved.