Open-Vocabulary 3D Perception Extends Beyond Fixed Category Sets
Three papers push 3D object detection and segmentation beyond closed category lists, but they target different tasks and sensor inputs.
Research Digest
Preprints and papers, rewritten in plain English.
New RL Preprints Focus on Stability, Generalization, and Evaluation
Five arXiv papers tackle persistent reliability problems in reinforcement learning, though they diagnose different failure modes.
Federated Learning Papers Tackle Efficiency, Heterogeneity, and Data Shift
Four new arXiv preprints show federated learning research moving beyond basic frameworks to address scale, uneven client capabilities, and unknown data distributions.
Forecasting Research Targets Physics, Retrieval, Coherence, and Architecture
Four new arXiv papers propose distinct improvements to time-series forecasting, from leveraging physical knowledge to aligning training objectives.
New Models Tackle Deformable Objects and Reusable Robot Skills
Four arXiv papers push robot manipulation beyond rigid parts, addressing deformable dynamics, asset generation, cable shaping, and skill reuse without demonstrations.
Policy Learning Papers Tackle Stability, Speed, and Visual Overfitting
Four arXiv preprints address distinct failure modes in policy learning, from experimental overfitting to visual domain shift.
New Research Tackles VLA Model Efficiency, Adaptation, and Data Scarcity
A wave of arXiv papers addresses the practical barriers to deploying vision-language-action models in robotics, from inference speed and fine-tuning costs to the need for more training data.
Gaussian Splatting Fixes for Reflections, PBR, Redundancy, and Online Mapping
New arXiv preprints address four distinct failure modes of 3D Gaussian Splatting: reflective surfaces, material decomposition, redundant Gaussians, and streaming RGB-D input.