Blog

MALLVI uses a closed‑loop multi‑agent framework that integrates LLM reasoning and vision feedback to improve robot generalization and task success.

RoboGene automates the generation of diverse, physically grounded robotic manipulation tasks to improve generalization and reduce manual dataset costs.

Vid2Sid uses paired sim/real video and vision‑language optimization to recover simulator physics parameters and improve sim‑to‑real robot transfer.

WildOS demonstrates how combining semantic reasoning with geometric navigation enables robots to explore and find objects efficiently using open‑vocabulary queries.

DreamDojo learns physical interactions from 44k hours of human videos to generate real‑time world models for robot planning and control.

Deep learning is unlocking new capabilities for robots to manipulate deformable objects like cloth — a major hurdle for automation in manufacturing and service robotics.

A CC BY Scientific Reports paper showing YOLO11n-seg crack segmentation at 3.6 ms/image (Tesla T4) with 78.8% precision.

A hybrid 1D-CNN + xLSTM pipeline for fatigue-state classification in CFRP using sensor-based SHM data, reporting 99.00% accuracy on the NASA-CFRP dataset.

An XR workflow that connects ROS2 to Unity3D for immersive debugging of navigation, SLAM, and sensor pipelines.

A robot-learning approach that adaptively chooses when to ask humans for feedback—and which feedback format—to reduce unsafe side effects efficiently.

Neurosim reports ~2700 FPS simulation for neuromorphic robot perception—potentially changing the economics of closed-loop iteration for robotics ML teams.

A pragmatic summary of a recent AEM electrolysis result, why it matters for scale-up, and a 10-day pilot plan engineering teams can actually run.