# Generalization Capability

R1-V is a project focused on enhancing the generalization capabilities of visual language models (VLMs). Using verified reward reinforcement learning (RLVR) technology, it significantly improves the generalization abilities of VLMs in visual counting tasks, particularly excelling in out-of-distribution (OOD) tests. The significance of this technology lies in its ability to efficiently optimize large-scale models at an extremely low cost (training costs as low as $2.62), offering new insights into the practical applications of visual language models. The project is based on improvements to existing VLM training methods, aiming to enhance model performance in complex visual tasks through innovative training strategies. Its open-source nature also makes it a vital resource for researchers and developers exploring and applying advanced VLM technologies.

SAMURAI is a visual object tracking model based on the Segment Anything Model 2 (SAM 2), specifically designed to handle the visual tracking task of fast-moving or self-occluding objects. It effectively predicts object motion and optimizes mask selection by introducing temporal motion cues and a motion-aware memory selection mechanism, achieving robust and accurate tracking without the need for retraining or fine-tuning. SAMURAI operates in real-time environments and demonstrates strong zero-shot performance across multiple benchmark datasets, proving its generalization capability without the need for fine-tuning. In evaluations, SAMURAI showed significant improvements in success rate and accuracy compared to existing trackers, with a 7.1% AUC increase on LaSOT-ext and a 3.5% AO increase on GOT-10k. Furthermore, compared to fully supervised methods on LaSOT, SAMURAI demonstrates competitive results, underscoring its robustness in complex tracking scenarios and potential practical value in dynamic environments.

ManiWAV is a research project aimed at learning robot manipulation skills from wild audio and visual data. It collects synchronized audio and visual feedback from human demonstrations and directly learns robot manipulation policies from these demonstrations through corresponding policy interfaces. The model demonstrates its system's capabilities through four contact-rich manipulation tasks, which require the robot to passively perceive contact events and patterns or actively perceive the material and state of object surfaces. Additionally, the system can generalize to unseen wild environments by learning from diverse wild human demonstrations.