WayTrace: A Map-to-Language-to-Trace Hierarchy for Embodied Urban Navigation
VLM 把网页地图路线拆成原子指令再画成图像轨迹，真机走完 1.37 km 城市路线
收录 发布
Vision-Language NavigationVisual NavigationLegged Robots Project page
FEIJOA: Task-Driven Grounding and Execution for Mobile Manipulation
按指令现场在原始扫描里找物体，再逐步核验动作；Spot 真机 50 个任务成功 35 个
收录 发布
Mobile ManipulationAgentsEmbodied Reasoning Project page
Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience
用相机位姿+投影骨架统一人和机器人的动作，人类视频越多预测越准
收录 发布
World ModelsHumanoidsMobile Manipulation CodeProject page
PadLoc: Training-Free Floor-Plan Localization by Sparse Geometric Matching
不用训练的平面图定位后端：墙体朝向+掩码相关匹配+多假设，胜过稠密网格滤波
收录 发布
Visual Localization Code
MoLRSI: A Recursive Self-Improving AI agent for Molecular Discovery
靠跨任务记忆改进搜索策略的科研 agent，分子筛选湿实验命中率近乎翻倍
收录 发布
AgentsSelf-Improvement Project page
Sparse Anchors for Visual Multi-Session Registration in Changing Environments
只信少量可靠锚点，再靠轨迹相对位姿，把变化工地的新一轮采集整条配进旧地图
收录 发布
Visual LocalizationPlace RecognitionLifelong SLAM
Asynchronous Robot Policies
慢 VLM 与快动作专家异步运行并用过期记忆训练，扰动下成功率大幅领先
收录 发布
VLARobot Learning
ParetoSaddler: Pareto-Guided Optimization of Agent Harnesses for Quality and Cost
自动改智能体外壳时，质量或花费任一项可信变好就放行，最后两项一起改善
收录 发布
Agents Project page
Caging Vision-Language Navigation for Legged Robots via Barrier-Certified Safety Architecture
给大模型导航加一层安全执行：VLM 只出目标，屏障约束的采样规划负责不撞
收录 发布
Vision-Language NavigationLegged RobotsHumanoids