Hybrid Particle-Voxel Dynamics for Multimodal 3D World Simulation
在3D体素潜空间里预测动作后果，任意视角渲染并附带触觉，跑到50Hz
收录 发布
World ModelsRobot Learning
每日论文雷达
每天挑出值得机器人 / 具身智能研究者看的几篇论文，用大白话讲清楚。Type 1 是能直接拿来用的工作，比如可以当基线、可以借鉴某个模块；Type 2 是方法比较有潜力、可能改变现有做法的论文。
最新
8 篇
单日页
GoalDex: Learning Image-Goal-Conditioned Policies for Dexterous In-Hand Manipulation
灵巧手不估位姿、只看目标图像转动物体，真机五物体成功率 80%
收录 发布
Robot LearningSim-to-Real Project page
SNAP: Semantic Navigation Among Pedestrians
把行人和地面语义画进同一张鸟瞰图，RL 策略边躲人边走人行道，零样本上四足
收录 发布
Legged RobotsSim-to-RealRobot Learning
TAMPAgent: Agentic Belief-Space Planning for Mobile Manipulation
让大模型当信念空间规划策略，工具认证约束后才动手，移动操作成功率升至77%
收录 发布
Mobile ManipulationAgentsEmbodied Reasoning Project page
MOST: Memory of Object States and Task Progress for Long-Horizon Mobile Manipulation
用物体状态和任务进度来监督策略记忆，多分钟厨房任务的最大进度约翻倍
收录 发布
Mobile ManipulationVLASpatial Memory
HeLM: Hierarchical Explicit Language Memory for Non-Markovian Manipulation
VLM 用文字记进度、事件触发才更新，带 VLA 做需要记忆的真机操作
收录 发布
VLAEmbodied ReasoningRobot Learning Project page
OpenNWM: A Generalist Navigation World Model with Latent Action Pretraining
用无动作视频学隐动作来预训练导航世界模型，预测更准、规划略胜 NWM
收录 发布
World ModelsVisual NavigationRobot Learning CodeProject page
Ray-SCR: Generalizing Scene Coordinate Regression Across Camera Models
给每个特征配上标定视线，一张全景地图让针孔、鱼眼、360°相机都能定位
收录 发布
Visual LocalizationMapping Project page
11 篇 单日页
WorldSketch: Sketching World Models for Zero-Shot Mobile Manipulation
VLM 只画任务相关的铰接草图，规划器出动作，厨房零样本成功率 91%
收录 发布
Mobile ManipulationArticulated ObjectsAgents
Do As I Can, Not As I Say: Grounding Language in Robotic Affordances
LLM 给技能打「有用分」、价值函数打「能成分」，相乘选动作，真机长程指令执行成功 74%
收录 发布
Mobile ManipulationEmbodied ReasoningFoundation Models Project page
Hierarchical World Models with Implicit Dynamics
把扩散噪声当高层动作搜子目标，长时程推T成功率从0.21升到0.77
收录 发布
World ModelsRobot Learning Project page
Behavior Graphs: Unsupervised Graph Structures for Interpreting Robot Deployment Data
把大量机器人回放聚类成行为图，免训练估进度且看失败更省力
收录 发布
Robot LearningVLA Project page
SPARK: Self-Correcting Embodied Memory via Physical Re-Observation
机器人自查记忆、飞回去重看再选择性改写，问答分从 0.31 升到 0.53
收录 发布
Spatial MemoryScene Graphs
Knowing Where to Look: Interactive Object Search via Spatial-Semantic Fusion
VLM 语义分与占据几何融成 3D 信念图，引导人形机器人开柜找物
收录 发布
Mobile ManipulationSpatial MemoryExploration Project page
Probing Spatial State Representations in Visual Foundation Models
用冻结特征探针测十种视觉基础模型的空间状态：看得清≠记得住≠用得上
收录 发布
Foundation ModelsSpatial MemoryWorld Models Project page
R^2Orch: Recursive Self-Improvement for Agent Orchestration via Self-Play within a Skill-Internalization Loop
编排器自己出题，边训练边撤掉提示里的技能，零技能部署下12个基准都领先
收录 发布
Agents Code
Learning to Live: A Step Toward Lifelong Embodied Intelligence
小模型流式门控决定何时记忆，平面图像记忆加自造工具，让机器人用一天经历答题
收录 发布
Spatial MemoryAgentsEmbodied Reasoning
Learning Agile Humanoid Dodgeball with Onboard Monocular Vision
G1 人形机器人只靠躯干单目相机躲球：用检测框历史估球位，三阶段课程训练全身闪避
收录 发布
HumanoidsLocomotionSim-to-Real
Auriga Technical Report: A Task-Agnostic Agent Harness for Zero-Shot Embodied Manipulation
靠“先预演、超差即停”的工具层，GPT零样本操作在RoboDojo排第一
收录 发布
AgentsEmbodied ReasoningMobile Manipulation CodeProject page
10 篇 单日页
StopWatch: Goal Grounding for Object-goal Navigation
物体导航常走到目标旁却不停；先让感知模型报出目标在否与距离，六个基准领先
收录 发布
Object-Goal NavigationVLABenchmarks Project page
Scaling Online Exploration for Robot Agents
冻结大模型的编程代理靠在线试错写出机器人程序，无示范在多数基准胜过示范训练策略
收录 发布
AgentsRobot LearningMobile Manipulation Project page
BeJEPA: A Goal-Conditioned Humanoid Behavior World Model via Unsupervised RL
一步动力学加后继特征训练人形全身控制，世界坐标全局误差降约六成
收录 发布
HumanoidsWorld ModelsRobot Learning Project page
FLORA: An Open Model Suite for Scaling Robotics Data Engines
一套开源模型同时给机器人视频打分段字幕和奖励，成本约为顶级 API 的 1/125
收录 发布
Robot LearningFoundation Models CodeProject page
Learning Force-Reactive Humanoid Loco-Manipulation
手持测力夹爪采示教加力位混合全身控制，人形真机成功率80%对26.7%
收录 发布
HumanoidsMobile ManipulationRobot Learning Project page
RESPLE2: Inertial-Informed Recursive Non-Uniform Spline Estimation for Multi-Sensor Odometry with Global Refinement
用 IMU 精确预测样条控制点的滤波里程计，剧烈跳跃下仍稳，代码开源
收录 发布
Legged RobotsMappingState Estimation Code
EgoContact: Evaluating Physical Intelligence with Egocentric Video
用2.25万条人类第一视角接触标注，考模型“该碰物体哪里”，并把机器人抓取成功率翻倍
收录 发布
BenchmarksRobot LearningFoundation Models Project page
RAVEN: A Unified Active SLAM System for Rapid Autonomous Visual Exploration with Aerial Robots under Localization Uncertainty
无人机主动 SLAM：用 λ 旋钮在探索速度与定位精度间取舍，真机全机载验证
收录 发布
ExplorationMappingVisual Localization
ScheduleGen: Demonstration Generation via Skill Scheduling for Multi-Arm Imitation Learning
把多臂协同写成 MILP 排程，5 条示教扩成千条无碰撞数据并零样本上真机
收录 发布
Mobile ManipulationRobot LearningSim-to-Real Project page
Learning Multi-Phase Robot Loco-Manipulation via Phase-Conditioned Policy Optimization
单个 PPO 策略按阶段给奖励和价值，真机完成人形搬箱与 Spot 推轮胎
收录 发布
Mobile ManipulationHumanoidsRobot Learning
没有符合条件的论文。