HAM-VLN: Harnessing Hierarchical Agentic Memory for Zero-Shot Vision-and-Language Navigation
HAM-VLN 为 zero-shot VLN agent 提供决策耦合的层次化记忆系统，通过世界图存储导航历史，结合动态门控检索，在无需训练的情况下超越现有方法
All tags
robotics
17 posts tagged with "robotics"
Reflex: Real-Time VLA Control through Streaming Inference
A streaming inference framework enabling real-time 50Hz control for Flow Matching Vision-Language-Action models via partitioned attention, AdaRMSNorm, and async pipeline
Scaling Mixture-of-Experts Video Pretraining for Embodied Intelligence
LingBot-Video — 首个开源大规模 MoE 视频基础模型，面向具身智能的单流扩散 Transformer，稀疏 MoE + 五阶段课程预训练 + 多维奖励 GRPO 后训练
V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning
Internet-scale JEPA video model pretrained on 1M+ hours of video, achieving SOTA understanding benchmarks and enabling zero-shot robotic planning via action-conditioned post-training with <62h robot data
V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised Learning
Next-gen JEPA model with dense predictive loss, deep self-supervision, and multi-modal tokenizers achieving SOTA on dense visual tasks (depth, segmentation, STA) while retaining global scene understanding
Advancing Open-source World Models
LingBot-World, an open-sourced world simulator stemming from video generation, featuring minute-level horizon, real-time interactivity, emergent memory, and three-stage evolutionary training pipeline
World Model for Robot Learning: A Comprehensive Survey
A comprehensive survey of world models from a robot-learning perspective, examining policy coupling, learned simulators for RL and evaluation, robotic video generation, and benchmarks/datasets across embodied applications
A Survey on Vision-Language-Action Models: An Action Tokenization Perspective
从动作 tokenization 视角统一审视 VLA 模型的综述。提出 VLA 模块 + 动作 token 的统一框架，将现有 VLA 模型归为 8 种动作 token 类型（语言描述/代码/可供性/轨迹/目标状态/隐表示/原始动作/推理），逐一分析每种类型的优势、局限与未来方向，并提出分层架构趋势与从 VLA 模型到 VLA 智能体的演进路径。
Vision-Language-Action Models for Robotics: A Review Towards Real-World Applications
面向真实世界部署的 VLA 全栈综述。系统梳理 VLA 三大挑战（数据稀缺 / 具身迁移 / 算力成本）、架构演进（CNN→Transformer→VLM→Diffusion/Flow→分层控制），并给出三类核心架构（sensorimotor 7 变体 / world model 3 模式 / affordance 3 模式）、模态处理（视觉/语言/动作/音频/触觉/3D）、训练范式（SL/SSL/RL + 预训练/后训练/梯度隔离/LoRA）、数据采集（遥操作/代理设备/人类视频）、公开数据集（Ego4D/OXE/DROID/AgiBot World）、机器人平台与评测基准（LIBERO/CALVIN/ManiSkill/SIMPLER/RoboArena）。最后给出面向实践者的 6 条建议与 8 个未来方向。
DIAL: Decoupling Intent and Action via Latent World Modeling
End-to-End VLA with Differentiable Latent Intent Bottleneck
OmniVLA-RL: A Vision-Language-Action Model with Spatial Understanding and Online RL
具有空间理解和在线强化学习的视觉-语言-动作模型
Wan-Move: Motion-controllable Video Generation via Latent Trajectory Guidance
通过潜在轨迹引导实现运动可控的视频生成
KV-Efficient VLA
A Method to Speed up Vision Language Models with RNN-Gated Chunked KV Cache
Qwen-VLA
统一的具身基础模型，通过DiT动作解码器将视觉-语言建模扩展到连续动作生成
VLA-Adapter: An Effective Paradigm for Tiny-Scale Vision-Language-Action Model
A lightweight bridging paradigm that enables SOTA-level VLA performance with only 0.5B parameters, without robotic data pre-training
Xiaomi-Robotics-0: An Open-Sourced Vision-Language-Action Model with Real-Time Execution
小米机器人团队发布的先进VLA模型，通过精心设计的训练方案和部署策略实现高性能实时执行
MobileVLA-R1: Reinforcing Vision-Language-Action for Mobile Robots
通过强化学习增强视觉-语言-动作框架，实现四足机器人的显式推理和连续控制