GlassFormer fuses radar and depth for real-time glass segmentation
A new model combines radar and depth data to help robots perceive transparent surfaces that standard cameras and depth sensors miss.
Research Digest
Preprints and papers, rewritten in plain English.
Geometry-Preserving Upper-Body Robot Motion Retargeting from Video
A new method maps upper-body human demonstrations from monocular video to robots while preserving geometric consistency, addressing the spatial mismatch between body and hand motion.
Decomposing Instructions Improves Language-Guided Robot Learning
A new method breaks task descriptions into finer-grained sub-instructions to better condition robot policies.
Turning Ordinary Videos Into Immersive Panoramic Experiences
A new approach aims to generate both wide-field-of-view video and matching spatial audio from standard footage.
Causal-EVC: Counterfactual Method for Emotional Video Captioning
A new arXiv preprint proposes breaking spurious emotional causality in video captioning via spatiotemporal grounding and counterfactual intervention.
Zero-Training Test-Time Adaptation via Information Geometry
A new arXiv note describes a test-time adaptation method that aligns feature spaces without iterative training.
Paper Explores In-Context Learning for Handwritten Text Recognition
An arXiv preprint examines whether in-context learning can be applied to handwritten text recognition, a technology used to digitize historical documents.
V-JEPA Policy: World-Action Models from Predictive Visual Latents
A new arXiv preprint proposes building world-action models on predictive visual latents rather than adapting pretrained video generators.
ResComEmb Paper Argues Single-Vector Compression Limits Multimodal Expressiveness
A new arXiv paper highlights how compressing multimodal inputs into a single vector can limit fine-grained expressiveness.
Cache-Aware Conv3D Lowering Boosts World Models on Edge Devices
A new approach to representing 3D convolutions in runtime systems could make generative world models more practical for embedded planning.
New Model Fuses Weather and Explainability for Robust Bus ETA
An adaptive meta-learning approach integrates weather and explainability to improve bus arrival time predictions under sparse and dynamic conditions.
Typed Interaction Features Aim to Forecast Open-Source Game Release Incidents
A new arXiv paper argues that conventional defect predictors miss the relational structure of open-source game development, proposing typed temporal interaction features instead.
Cross-Material Transfer Improves Core-Loss Prediction for Real-World Waveforms
A new method adapts core-loss predictions from lab-friendly waveforms to the trapezoidal and PWM shapes seen in actual power converters.
New benchmark tests LLM agents on causal discovery under budget
A new SCM-grounded environment evaluates whether LLM agents can recover causal graphs from observations and budget-constrained hard interventions.
Satellite Onboard Learning Gets Resource-Aware Federated Method
A new framework adapts federated mixture-of-experts with adaptive pruning for low-Earth-orbit satellites, where conventional federated learning falls short.
MPC for Knee Rehab Exoskeletons Treats Spasms as Disturbances
A new control framework for knee rehabilitation exoskeletons frames safe interaction as a disturbance-rejection problem, absorbing spasms and voluntary effort.
Testing Physics-Inspired Neural Networks on Dissipative Systems
A new preprint examines whether Lagrangian and Hamiltonian neural networks can handle systems with explicit time dependence and dissipation.
Predicting OOD Generalization from Weights Alone
A new arXiv preprint asks whether circuit alignment in a trained model's weights can forecast out-of-distribution generalization without any target-domain data.