#VLM
4 related articles
VLM Agents Self-Evolve in Latent Worlds: A Deep Dive into the Dwell Framework
The Dwell framework reframes VLMs as recurrent latent transition operators, letting agents predict and act in latent space, achieving the strongest overall performance across five interactive visual environments while greatly reducing imagination computation.
StopWatch: How Goal Grounding Breaks the Success-Rate Bottleneck in Object-Goal Navigation
StopWatch identifies goal grounding as the main failure cause in object-goal navigation, achieves SOTA on six benchmarks via explicit grounding, fixes benchmark annotation flaws, and shows perception—not policy learning—is the ceiling.
HexaShell Lets Robots Write Their Own Code to Fix Grasping Failures
HexaShell framework lets vision-language models write and debug robot code directly. Robots generate corrective scripts after failed grasps, completing tasks in under ten seconds. RLBench pick-and-place reaches 78% success within three attempts, revealing VLM programming capabilities and three bottlenecks: depth estimation, temporal errors, and token limits.
TutorialsMiniMind-V: A Complete Tutorial to Train a 65M Parameter Vision Multimodal LLM from Scratch in 2 Hours
MiniMind-V is a lightweight open-source VLM project that trains a 65M parameter vision multimodal model from scratch in just 2 hours. Learn its architecture, training pipeline, and educational value.