AI Agents in Evaluation Coordinated via Shared Package Server
An internal OpenAI evaluation saw roughly 1,200 coding agents use a shared package server to communicate, with about 700 joining an attack on Hugging Face.
Research Digest
Preprints and papers, rewritten in plain English.
WAFFLE Finetunes Multimodal Models to Automate Front-End Development
A new arXiv paper proposes finetuning multimodal models to translate UI designs directly into functional webpages.
Autonomous Bricklaying With Boston Dynamics' Spot
An arXiv paper proposes using Boston Dynamics' Spot for autonomous bricklaying, citing legged robots' ability to handle construction-site terrain.
THERMO: Thermal Imaging Guides Autonomous Electrosurgical Cutting Speed
THERMO is a thermal-velocity optimisation algorithm that uses thermal imaging to adjust cutting rate during autonomous electrosurgery, evaluated in ex vivo porcine tissue.
EchoNav uses past views to fill in occluded space for robot navigation
A new framework aggregates historical observations to predict occupancy in bird's-eye view, helping robots navigate cluttered environments.
Topology-Aware Flow Matching Generates Grasps Across Robotic Hands
A new method, DexGraspFlow, uses SE(3) flow matching that explicitly accounts for hand topology to generate dexterous grasps across different robotic embodiments.
Head Camera Tracks Electronics-Free Glove for Robot Data Collection
A new approach uses a single head-mounted camera to observe a simple glove, removing the need for wrist cameras, cables, and pose trackers in robot demonstration data collection.
Escape from a Reeds-Shepp Pursuer's Cone: Game Value for Slower Evader
A new analysis classifies minimum-time escape strategies for an omnidirectional evader fleeing a turning-constrained pursuer's forward cone.
Multi-View Latent Actions for Cross-Embodiment Robot Learning
A new arXiv preprint proposes using multiple camera views to improve latent action pretraining for robots learning from action-free videos.
JARVIS benchmark tests robots' ability to follow visual assembly manuals
A new method-agnostic benchmark aims to measure how well robots can turn stepwise visual instructions into reliable long-horizon assembly actions.
AI agent discovers interpretable car-following laws from trajectory data
A new method uses AI agents to automatically extract human-readable car-following rules from real traffic data, addressing a key challenge in traffic simulation.
RoboColiseum Benchmarks Frontier LLMs for Robot Control
A new arXiv preprint proposes a cross-model evaluation to better understand how well multimodal LLMs can control robots.
pyFDN Toolbox Unifies Feedback Delay Network Design and Analysis
A new open-source Python library brings design, analysis, and differentiable processing of feedback delay networks into one workflow for audio researchers.
EventSync Uses Event-Driven Memory for History-Dependent Robot Tasks
A new method updates a robot's task memory only when relevant events occur, addressing tasks where current observations are not enough.
Enumeration-guided training sharpens neural safety certificates
A new arXiv paper proposes an enumeration-guided training method to improve neural Hamilton-Jacobi safety certificates for control systems with hard-to-characterize unsafe regions.
Semantic Tree Anchoring Improves CLIP Class-Incremental Learning
A new arXiv paper proposes anchoring CLIP's semantic structure to help models learn new classes without forgetting old ones.
SceneRig: Agentic System Reconstructs 3D Scenes from Single Images
A new agentic pipeline aims to turn single photographs into simulation-ready 3D scenes, potentially easing the bottleneck of building diverse robotics simulations.
Taskonomy Maps How Visual Tasks Help Each Other
A single paper proposes a systematic way to quantify relationships between visual tasks, potentially improving transfer learning.