今天 world model
VLA 線 · DEEP DIVE ARCHIVEVLA 线 · DEEP DIVE ARCHIVE
VLA 深度追蹤VLA 深度追踪
Vision-Language-Action：讓機器人看→想→做的端到端模型Vision-Language-Action：让机器人看→想→做的端到端模型
VLA-Handbook 庫VLA-Handbook 库 theory/ 目錄theory/ 目录 Evo-SOTA 排行榜Evo-SOTA 排行榜 →
論文全景圖论文全景图 130+
206 理論文章理论文章 · 30 SOTA 記錄SOTA 记录
METHOD FAMILY TRENDS
high
METHOD FAMILY TRENDS
▼ 3 declining 15 families · 241 papers covered
FAMILY MOM 7d 14d 30d Δ7d Δ14d Δ30d CHART ST
Lang. Grounding 72 120 264 1.31x 0.73x 1.14x ●
RL Fine-tuning 31 43 95 1.29x 0.69x 1.13x ●
Multi-Task 23 38 69 0.95x 0.84x 0.82x ●
Flow Matching 19 27 56 0.79x 0.60x 0.66x ●
Long Horizon 18 25 69 0.75x 0.48x 0.82x ●
Human-Robot 16 21 61 0.66x 0.43x 0.72x ●
World Model 15 26 87 0.62x 0.36x 0.73x ●
Tactile 12 14 41 0.50x 0.31x 0.49x ●
Dexterous Hand 11 22 46 0.46x 0.49x 0.55x ●
Cross-Embodiment 7 11 22 0.29x 0.24x 0.26x ●
Mobile Manip. 6 9 32 0.25x 0.20x 0.38x ●
Diffusion Policy 3 4 20 0.12x 0.09x 0.24x ●
Instr. Tuning 3 3 7 0.12x 0.07x 0.08x ●
Sim-to-Real 3 7 23 0.12x 0.15x 0.27x ●
3D Repr. 2 3 7 0.08x 0.07x 0.08x ●
COMPETITION PAIRS 6 matchups · hover for details
VLA vs WAM
Paradigm war: end-to-end action prediction vs world model planning
Lang. Ground. vs World Model
9.6% · x1.31 ratio 4.80 2.0% · x0.62
Lang. Grounding
Language Grounding: connecting natural language instructions to robot actions; vision-language-action alignment
World Model
World Model: learned environment simulator (Dreamer, UniSim); enables planning via imagination without real-world interaction
Why they compete
The central paradigm war in embodied AI. VLA (Vision-Language-Action) maps observations directly to actions end-to-end — simple, scalable, but needs massive data and generalizes poorly. WAM (World-Action Model) first learns how the world works, then plans actions through mental simulation — better generalization and data efficiency, but world models are often inaccurate. The boundary is blurring: Pi0.5 uses flow matching (generative, WAM-like), GR00T adds video prediction. The winner likely is a hybrid.
ACTION HEAD ROUTE
Continuous action generation: denoising vs optimal transport
Diffusion Pol. vs Flow Matching
0.4% · x0.12 ratio 0.16 2.5% · x0.79
Diffusion Policy
Diffusion Policy: iterative denoising process (DDPM) to generate continuous robot actions; strong on multi-modal action distributions
Flow Matching
Flow Matching: optimal-transport-based generative model (e.g. Pi0); faster inference than diffusion with comparable quality
Why they compete
Both generate continuous actions from the same VLA backbone but take different mathematical routes: diffusion iteratively denoises random noise into actions (slow, expressive), while flow matching uses optimal transport for a direct trajectory (fast, efficient). If flow matching matches diffusion quality, it could replace it as the default action head.
POST-TRAINING ROUTE
Model adaptation: supervised tuning vs reward optimization
Instr. Tuning vs RL Fine-tune
0.4% · x0.12 ratio 0.10 4.1% · x1.29
Instr. Tuning
Instruction Tuning: supervised fine-tuning (SFT) on language-action pairs; simpler but limited to offline data distribution
RL Fine-tuning
RL Fine-tuning: post-training with PPO/DPO/GRPO reward signals; enables online improvement beyond demonstration data
Why they compete
After pretraining a VLA, two competing strategies exist: SFT directly imitates expert demonstrations (simple, stable), while RL fine-tuning (GRPO/DPO) optimizes a reward signal to go beyond the demonstration distribution. RL can discover novel strategies but is harder to stabilize.
LEARNING SIGNAL
Training paradigm: imagination-based vs reward-based
World Model vs RL Fine-tune
2.0% · x0.62 ratio 0.48 4.1% · x1.29
World Model
World Model: learned environment simulator (Dreamer, UniSim); enables planning via imagination without real-world interaction
RL Fine-tuning
RL Fine-tuning: post-training with PPO/DPO/GRPO reward signals; enables online improvement beyond demonstration data
Why they compete
World models learn by predicting the future (imagination-based planning), while RL learns from reward feedback. If world models become accurate enough, they could reduce the need for expensive real-world RL exploration.
MANIPULATION SENSING
Manipulation approach: tactile feedback vs dexterous control
Tactile vs Dext. Hand
1.6% · x0.50 ratio 1.09 1.5% · x0.46
Tactile
Tactile Sensing: force/torque and GelSight contact sensors; provides direct manipulation feedback for delicate tasks
Dexterous Hand
Dexterous Hand: multi-finger manipulation control; achieves fine-grained object interaction without dedicated sensors
Why they compete
Two approaches to dexterous manipulation: tactile sensing adds explicit touch feedback (hardware cost, rich signal), while dexterous hand control relies on proprioception and vision alone (simpler hardware, harder control). The winner depends on sensor cost-to-performance ratio.
TRANSFER APPROACH
Domain bridging: simulation transfer vs cross-embodiment
Sim2Real vs Cross-Embod.
0.4% · x0.12 ratio 0.43 0.9% · x0.29
Sim-to-Real
Sim-to-Real: train in simulation, deploy on real hardware; uses domain randomization to bridge the reality gap
Cross-Embodiment
Cross-Embodiment: transfer policies across different robot morphologies; aims for universal robot foundation models
Why they compete
Sim-to-Real trains one robot in simulation then transfers (cheap data, reality gap risk), while Cross-Embodiment trains across multiple real robots directly (expensive data, natural generalization). The approaches represent different bets on where generalization should happen.
EMERGING SIGNALS
1 signals
EMERGING SIGNALS
TERM COUNT AGE VELOCITY STATUS SAMPLE
agentic robotics
5 6d ~ x2.0 RISING AgenticRoboRL: Bringing Coding-Agent Post-Training... TOP INSTITUTIONS
20 active / 30d
TOP INSTITUTIONS
INSTITUTION TOTAL BEST LAST SEEN ACTIVITY
1 Stanford 9 10-24
2 LIBERO Team 9 10-18
3 CMU 7 10-24
Institute for AI Industry Research (AIR), Tsinghua University 5 10-04
KAIST 4 10-04
MIT 3 09-27
University of Maryland, College Park 3 10-01
清华 3 10-14
Samsung Robotics eXperience 2 09-27
KTH Royal Institude of Technology 2 09-27
China Mobile (Hangzhou) Information Technology Co., Ltd 2 10-02
AIST 2 10-04
Southern University of Science and Technology, Shenzhen, China 2 10-04
Shanghai Jiao Tong University 12 10-24
Harbin Institute of Technology 5 10-01
University of Southern California 3 10-02
Beihang University 2 10-01
Carnegie Mellon University 2 10-24
Microsoft Research Asia 2 09-29
Wu 2 10-21
📐 理論文章庫📐 理论文章库
206 篇篇 查看 GitHub 全庫查看 GitHub 全库 →
最近 2 週最近 2 周 91 篇篇
今天 rl
什么时候停止惩罚动作？面向 VLA 强化学习的 Martingale GRPO 拆解 (When to Stop Penalizing Actions? Martingale GRPO for VLA Reinforcement Learning)
在 GitHub 閱讀在 GitHub 阅读 昨天 world model
DRILL：训练世界模型以改进策略，而非预测像素 (DRILL: Training World Models to Improve Policies, Not Predict Pixels)
在 GitHub 閱讀在 GitHub 阅读 昨天 diffusion flow
一步 Euler 够不够？Flow-Matching VLA 的步数等价性实证 (When One Step Is Enough: An Empirical Study of Euler Integration in Flow-Matching Policies)
在 GitHub 閱讀在 GitHub 阅读 昨天 vla core
FastHarness：用经验调度 GPT-6 复核，降低 VLA 执行监控成本
在 GitHub 閱讀在 GitHub 阅读 昨天 vla core
FocusVLA：未来意图引导的多视角动态路由 (FocusVLA: Future-Intent Guided Dynamic View Routing for Vision-Language-Action Models)
在 GitHub 閱讀在 GitHub 阅读 2 天前 vla core
Progress Probes P²：用 VLA 内部激活做轻量运行时监控 (Progress Probes P²: Lightweight Runtime Monitoring for Vision-Language-Action Models)
在 GitHub 閱讀在 GitHub 阅读 2 天前 vla core
Echo：VLA 动作与任务状态信号的耦合去噪 (Echo: Coupled Action and Task Status Signal Denoising for Vision Language Action Models)
在 GitHub 閱讀在 GitHub 阅读 3 天前 frontier
UniMem：把视觉与文本记忆统一进 VLA 的非马尔可夫任务框架 (UniMem: Unifying Visual and Textual Memory in VLAs for Non-Markovian Tasks)
在 GitHub 閱讀在 GitHub 阅读 3 天前 rl
超越示教：面向 VLA 的强化学习仿真—真实协同训练 (Beyond Imitation: Reinforcement Learning-Based Sim-Real Co-Training for VLA Models)
在 GitHub 閱讀在 GitHub 阅读 3 天前 frontier
从非机器人合成数据中学习机器人行为先验 (Learning Robot Behavior Priors from Non-Robotic Synthetic Data)
在 GitHub 閱讀在 GitHub 阅读 3 天前 frontier
前史记忆：人形 Loco-Manipulation 的对象事件关键帧记忆 (Memory before Goal: Role of Pre-History Memory for Humanoid Loco-Manipulation)
在 GitHub 閱讀在 GitHub 阅读 3 天前 planning
感知运动对齐：具身策略成功的柏拉图式代理 (Sensorimotor Alignment: A Platonic Proxy for Embodied Policy Success)
在 GitHub 閱讀在 GitHub 阅读 4 天前 world model
UniWALM：把语言变成世界-动作模型的工作记忆 (UniWALM: Unified World Action Language Model for Embodied Robot Control)
在 GitHub 閱讀在 GitHub 阅读 4 天前 frontier
DEFLECT：用时间反事实偏好学习训练抗延迟的异步 VLA
在 GitHub 閱讀在 GitHub 阅读 4 天前 world model
μ0：可扩展的 3D 交互痕迹世界模型（$\mu_0$: A Scalable 3D Interaction-Trace World Model）
在 GitHub 閱讀在 GitHub 阅读 4 天前 planning
Safe-Night VLA：给 VLA 加一路热成像，再用 CBF-QP 保住执行安全 (Safe-Night VLA: Seeing the Unseen via Thermal-Perceptive Vision-Language-Action Models for Safety-Critical Manipulation)
在 GitHub 閱讀在 GitHub 阅读 4 天前 tactile
物理画布：面向机器人策略评估与学习的视觉-触觉世界模型（Physical Canvas: A Vision Tactile World Model for Robot Policy Evaluation and Learning）
在 GitHub 閱讀在 GitHub 阅读 5 天前 vla core
异步 VLA 如何让无人机在推理延迟中继续飞：AeroDASH 精读 (AeroDASH: Asynchronous Vision-Language-Action Control for Continuous UAV Flight)
在 GitHub 閱讀在 GitHub 阅读 5 天前 deployment
T(R,O) Grasp++：杂乱场景下的跨本体灵巧抓取规模化 (T(R,O) Grasp++: Cross-Embodiment Dexterous Grasping in Cluttered Environments at Scale)
在 GitHub 閱讀在 GitHub 阅读 5 天前 world model
Vector-Quantized World Model：用控制器可达性做离散 latent 规划的机器人操作框架 (Vector-Quantized World Model: Latent Feedback Motion Planning for Robot Manipulation)
在 GitHub 閱讀在 GitHub 阅读 5 天前 vla core
不是所有特征都生而平等：VLA 的机制可解释性研究 (Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Models)
在 GitHub 閱讀在 GitHub 阅读 6 天前 planning
NativeMEM：把长程视觉历史压进 VLA 原生 Token 空间 (Native Memory Compression for Long-Horizon Robotic Manipulation)
在 GitHub 閱讀在 GitHub 阅读 6 天前 frontier
HexaShell：用薄封装探测 VLM 编码 Agent 作为机器人编码 Agent 的上限 (HexaShell: A Thin Harness for Exploring the Limits of VLM Agents as Robot Coding Agents)
在 GitHub 閱讀在 GitHub 阅读 6 天前 frontier
Vision-Language-Delta-Action：把紧凑 Delta 作为可操作的状态转移接口 (Vision-Language-Delta-Action: Learning Compact Deltas as Actionable Transition Interfaces for Robotic Manipulation)
在 GitHub 閱讀在 GitHub 阅读 6 天前 diffusion flow
CFG-Policy：用 Classifier-free Guidance 统一扩散策略的采样、探索与 RL 微调 (CFG-Policy: Unified Diffusion Policy Learning Framework via Classifier-free Guidance)
在 GitHub 閱讀在 GitHub 阅读 7 天前 vla core
LTM：面向 VLA 的潜转移上下文记忆（Latent Transition Based In-Context Memory for Vision-Language-Action Models）
在 GitHub 閱讀在 GitHub 阅读 7 天前 vla core
FastVLA：任务条件化与时间自适应的 VLA 视觉 Token 压缩（FastVLA: Task-Conditioned and Temporally Adaptive Visual Token Compression for Efficient Vision-Language-Action Models）
在 GitHub 閱讀在 GitHub 阅读 7 天前 frontier
REKI：把记忆注册进 VLA（REKI: Registering Knowledge in VLAs）
在 GitHub 閱讀在 GitHub 阅读 8 天前 frontier
把历史写进权重：流式视频的长期记忆内化 (Memory as Weights: Internalizing Long-Term History for Streaming Videos)
在 GitHub 閱讀在 GitHub 阅读 8 天前 frontier
测试时计算重塑人形行为基座模型 (Test-Time Compute for Humanoid Behavior Foundation Models)
在 GitHub 閱讀在 GitHub 阅读 8 天前 planning
Mem4Act：为长时程机器人操作学习有效潜记忆（Mem4Act: Learning Effective Latent Memory for Long-Horizon Robotic Manipulation）
在 GitHub 閱讀在 GitHub 阅读 8 天前 world model
记住该记住的：长程机器人世界模型的上下文调度 (Remembering What Matters for Long-Horizon Robotic World Models)
在 GitHub 閱讀在 GitHub 阅读 8 天前 vla core
只思考一次的 VLA：Fast-VLA 的任务记忆与闭环解耦 (Fast-VLA: You Only Think Once for Robot Manipulation)
在 GitHub 閱讀在 GitHub 阅读 8 天前 tactile
N₀-VTLA：把触觉变成“未来接触”的 latent 条件
在 GitHub 閱讀在 GitHub 阅读 8 天前 tactile
VLA-Touch：用双层触觉反馈增强 VLA / Enhancing Vision-Language-Action Models with Dual-Level Tactile Feedback
在 GitHub 閱讀在 GitHub 阅读 9 天前 tactile
Spark-0：面向机器人灵巧操作的空间对齐触觉预训练 (Spark-0: Scaling Spatially Aligned Tactile Pretraining for Robot Dexterous Manipulation)
在 GitHub 閱讀在 GitHub 阅读 9 天前 vla core
机器人策略该学什么预测？重审 VLA 学习的监督目标 (What Should Robot Policies Learn to Predict? Rethinking Supervision Targets in Vision-Language-Action Learning)
在 GitHub 閱讀在 GitHub 阅读 9 天前 foundation
VLA 真的在“听话”吗？How-level 指令跟随诊断 (Do VLAs Really Follow Instructions? Diagnosing How-Level Instruction Following in Vision-Language-Action Models)
在 GitHub 閱讀在 GitHub 阅读 9 天前 frontier
未来一致才提交：面向机器人操作的结果感知自适应动作分块（Commit While Futures Agree: Consequence-Aware Adaptive Action Chunking for Robot Manipulation）
在 GitHub 閱讀在 GitHub 阅读 9 天前 planning
EgoLAP：用语言-动作推理从第一视角人类数据中学习 (EgoLAP: Learning from Egocentric Human Data through Language-Action Reasoning)
在 GitHub 閱讀在 GitHub 阅读 9 天前 frontier
部署中协同演化机器人编排器与策略 (Co-Evolving Robot Orchestrators and Policies through Deployment)
在 GitHub 閱讀在 GitHub 阅读 9 天前 vla core
从人类视频到机器人操作的跨本体因果迁移 (Cross-Embodiment Causal Transfer from Human Videos to Robot Manipulation)
在 GitHub 閱讀在 GitHub 阅读 9 天前 frontier
Beyond Reconstruction：动作 Tokenization 到底需要什么？(Beyond Reconstruction: What Matters in Action Tokenization for Robot Policies?)
在 GitHub 閱讀在 GitHub 阅读 9 天前 tactile
工作空间级等变视觉-触觉扩散策略：VISTA (Equivariant Visual-Tactile Diffusion Policy for Contact-Rich Manipulation)
在 GitHub 閱讀在 GitHub 阅读 9 天前 tactile
TouchScale：500 小时人类视觉-触觉数据 (TouchScale: 500 Hours of Human Vision and Touch for Visual-Tactile Learning)
在 GitHub 閱讀在 GitHub 阅读 9 天前 foundation
多教师在线策略蒸馏：把域专用 Flow-Based VLA 合并进共享学生 (Multi-Teacher On-Policy Distillation for Flow-Based Vision-Language-Action Models)
在 GitHub 閱讀在 GitHub 阅读 9 天前 vla core
Juno：驯服预测隐变量用于 VLA 模型 (Juno: Taming Predictive Latents for Vision-Language-Action Models)
在 GitHub 閱讀在 GitHub 阅读 9 天前 rl
多种成功路径：面向 VLA 泛化的多样性驱动 RL 微调 (Many Ways to Succeed: Diversity-Driven RL Fine-Tuning for VLA Generalization)
在 GitHub 閱讀在 GitHub 阅读 9 天前 vla core
改写先行：刻画并缓解 VLA 的语言敏感性 (Rephrase Before You Act: Characterizing and Mitigating Language Sensitivity in Vision-Language-Action Models)
在 GitHub 閱讀在 GitHub 阅读 9 天前 tactile
TACROSS：跨异构触觉传感器的高效低成本人体触觉采集系统 (TACROSS: An Efficient and Low-Cost Scalable Human Touch System Across Heterogeneous Tactile Sensors for Dexterous Robot Learning)
在 GitHub 閱讀在 GitHub 阅读 9 天前 vla core
REACT：滚动去噪与双重解耦的 VLA 反应式机器人控制（REACT: Rolling Denoising and Dual Decoupling for Reactive Robot Control with VLA Models）
在 GitHub 閱讀在 GitHub 阅读 9 天前 tactile
OpenViTac：统一仿真-真实框架下的视触觉策略学习与基准评测 (OpenViTac: Learning and Benchmarking Visuo-Tactile Policies in a Unified Sim-and-Real Framework)
在 GitHub 閱讀在 GitHub 阅读 9 天前 world model
PLaW-VLA：面向视觉-语言-动作策略的预测性潜在世界建模 (Predictive Latent World Modeling for Vision-Language-Action Policies)
在 GitHub 閱讀在 GitHub 阅读 9 天前 planning
ARC：面向机器人基础模型的推理后训练配方（Arc: A Reasoning Recipe for Robot Foundation Models）
在 GitHub 閱讀在 GitHub 阅读 9 天前 vla core
ValueFormer：阶段感知价值标签，让 VLA 的「静默失败」可被在线读出 (ValueFormer: Stage-Aware Value Labels for Online Failure Detection and VLA Post-Training)
在 GitHub 閱讀在 GitHub 阅读 9 天前 vla core
VersaCamVLA：面向机器人操作的相机可配置 VLA 策略 (VersaCamVLA: Camera-Configurable VLA Policies for Robotic Manipulation)
在 GitHub 閱讀在 GitHub 阅读 9 天前 frontier
前向动力学模型：把世界模型变成动作特征的后置监督 (Forward Dynamics Models)
在 GitHub 閱讀在 GitHub 阅读 9 天前 vla core
双流记忆：把 VLA 的“执行状态”和“视觉证据”拆开 (Dual-Stream Memory for Vision-Language-Action Policies)
在 GitHub 閱讀在 GitHub 阅读 9 天前 frontier
证据引导的反事实学习：长尾机器人策略的头部动作偏置诊断与缓解 (Evidence-Guided Counterfactual Learning for Long-Tailed Robotic Policies)
在 GitHub 閱讀在 GitHub 阅读 9 天前 vla core
Splat2Act：从捕获场景到个性化 VLA 操作（Splat2Act: From Captured Scenes to Personalized VLA Manipulation）
在 GitHub 閱讀在 GitHub 阅读 9 天前 vla core
噪声空间学习：冻结 flow-matching VLA 的轻量适配 (Noise Space Learning for VLA Adaptation)
在 GitHub 閱讀在 GitHub 阅读 9 天前 vla core
具身中间训练：用数据邻近度弥合 VLM 与 VLA 之间的鸿沟 (EmbodiedMidtrain: Bridging the Gap between Vision-Language Models and Vision-Language-Action Models via Mid-training)
在 GitHub 閱讀在 GitHub 阅读 9 天前 vla core
Evo-World：用光流监督未来运动预测的轻量 VLA (Evo-World: A Lightweight Vision-Language-Action Model with Future Motion Prediction via Optical-Flow Supervision)
在 GitHub 閱讀在 GitHub 阅读 9 天前 deployment
以腕部视角人类演示构建跨灵巧手可复用数据接口 (Human2Dex: Dexterous Robot Learning from Wrist-Centric Human Demonstrations)
在 GitHub 閱讀在 GitHub 阅读 9 天前 vla core
AFCE：为 VLA 学习“未来操作目标”的效果监督 (Beyond Actions: Learning Future Operation Targets for Vision-Language-Action Models)
在 GitHub 閱讀在 GitHub 阅读 9 天前 rl
AlphaNIST：从二值任务奖励中涌现的装配技能 (AlphaNIST: Emergent Assembly Skills from Binary Task Rewards)
在 GitHub 閱讀在 GitHub 阅读 9 天前 vla core
报警值不值得响应：VLA 运行时监控器的反事实评估 (RESCUE: Is This Alarm Worth Acting On? Counterfactual Evaluation of VLA Runtime Monitors)
在 GitHub 閱讀在 GitHub 阅读 9 天前 planning
检索锚定的鲁棒策略微调：RARA 把“相关预训练经验”变成表征锚点 (RARA: Robust Policy Fine-Tuning through Retrieval-Anchored Representation Alignment for Robotic Manipulation)
在 GitHub 閱讀在 GitHub 阅读 9 天前 deployment
JitterVLA：面向 VLA 驱动机器人操作运动质量的回放基准 (JitterVLA: A Replay Benchmark for Motion Quality in VLA-Driven Human-Robot Manipulation)
在 GitHub 閱讀在 GitHub 阅读 9 天前 deployment
AutoBio+：真实生物实验室中动态实验现象的灵巧操作基准 (AutoBio+: Benchmarking Dexterous Manipulation with Dynamic Experimental Phenomena in Realistic Laboratories)
在 GitHub 閱讀在 GitHub 阅读 9 天前 foundation
LAPT-Scale：用人类视频连续隐动作预训练扩展具身策略 (LAPT-Scale: Scaling Embodied Policies with Latent Action Pre-Training on Human Videos)
在 GitHub 閱讀在 GitHub 阅读 9 天前 world model
SmolWAM：面向边缘设备的隐空间世界-动作模型（SmolWAM: A Latent World-Action Model for Robotic Manipulation on Edge Devices）
在 GitHub 閱讀在 GitHub 阅读 10 天前 frontier
免整理记忆：为机器人策略学习多流潜在记忆 (Memory Without Curation: Learning Multi-Stream Latent Memory for Robot Policies)
在 GitHub 閱讀在 GitHub 阅读 10 天前 diffusion flow
MOSS：面向机器人流匹配策略的模式导向源分布塑造 (Mode-Oriented Source Shaping for Robot Flow Matching Policies)
在 GitHub 閱讀在 GitHub 阅读 10 天前 world model
VeriVLA：基于潜世界模型的自适应测试时验证，用于流式 VLA (VeriVLA: Adaptive Test-Time Verification for Flow-Based VLAs with Latent World Models)
在 GitHub 閱讀在 GitHub 阅读 10 天前 world model
TS-WAM：时间潜在动作与世界-动作耦合的视觉语言机器人控制 (TS-WAM: Temporal Latent Actions and World-Action Coupling for Vision-Language Robot Control)
在 GitHub 閱讀在 GitHub 阅读 11 天前 vla core
精密強迫：讓通用 VLA 策略跨過亞毫米容差門檻 (PRISM: Precision Forcing for Generalist Vision-Language-Action Policies)
在 GitHub 閱讀在 GitHub 阅读 11 天前 vla core
把測試時計算蒸餾進策略：Verify, Act, Distill (Turning Test-Time Compute into Test-Time Adaptation for Vision-Language-Action Models)
在 GitHub 閱讀在 GitHub 阅读 11 天前 world model
因果动作增益世界动作模型 (CAG-WAM: Causal Action-Gain World Action Models for Robot Manipulation)
在 GitHub 閱讀在 GitHub 阅读 11 天前 tactile
DEFT-VLA：面向接触丰富操作的分阶段力觉专精 (DEFT-VLA: Stage-Localized Force Specialization for Contact-Rich Robotic Manipulation)
在 GitHub 閱讀在 GitHub 阅读 11 天前 vla core
通过质量多样性环境生成对 VLA 做红队评测 (Red-Teaming of Vision-Language-Action Models via Quality Diversity Environment Generation)
在 GitHub 閱讀在 GitHub 阅读 11 天前 vla core
把测试时计算蒸馏进策略：Verify, Act, Distill (Turning Test-Time Compute into Test-Time Adaptation for Vision-Language-Action Models)
在 GitHub 閱讀在 GitHub 阅读 12 天前 frontier
USER：面向真实世界在线策略学习的统一可扩展系统 (USER: A Unified and Extensible System for Real-World Online Policy Learning in Embodied AI)
在 GitHub 閱讀在 GitHub 阅读 12 天前 planning
BrickCraft：用「情境说明书」把符号装配计划落到实时视觉上 (Visuomotor Skill Composition with Situated Manual Guidance for Long-Horizon Interlocking Brick Assembly)
在 GitHub 閱讀在 GitHub 阅读 12 天前 deployment
人类-机器人对齐表征学习：用配对桥接把人类视频变成 VLA 监督 (HARP-VLA: Human-Robot Aligned Representation Learning for Vision-Language-Action Model)
在 GitHub 閱讀在 GitHub 阅读 13 天前 frontier
显式动作空间接地：让预训练 VLM 原生成为机械臂动作模型 (VLMs are Native Action Models: Unleashing the Manipulation Capabilities of VLMs with Explicit Action-Space Grounding)
在 GitHub 閱讀在 GitHub 阅读 13 天前 vla core
OGPO：面向可恢复 VLA 的离线目标条件策略优化 (Offline Goal-conditioned Policy Optimization for Recoverable Vision-Language-Action Models)
在 GitHub 閱讀在 GitHub 阅读 13 天前 world model
Grounded-Dreamer：以物理錨定的世界模型合成資料訓練機器人操作策略 (Grounded-Dreamer: Robot Learning from World Model Synthetic Data with Physical Grounding)
在 GitHub 閱讀在 GitHub 阅读 13 天前 planning
先想象未来，再内化为直觉：IG-VLA 的高效时空推理 (Imagine the Future, Internalize the Gist: Efficient VLA Reasoning via Internalized Spatiotemporal Imagination)
在 GitHub 閱讀在 GitHub 阅读 13 天前 frontier
用 Feynman–Kac 后验采样给预训练策略上约束 (BayesFP: Posterior Estimation for Flow-Based Policies via Feynman-Kac Sampling)
在 GitHub 閱讀在 GitHub 阅读更多文章 · 全部在 GitHub更多文章 · 全部在 GitHub 115 篇篇
🏛️ VLA Core · 34 篇篇
检测并抑制：面向 VLA 对抗补丁的机制级防御 (Detect and Suppress: A Mechanistic Defense against Adversarial Patches in VLA Models) 分而记之：面向长时序 VLA 策略的递归动作相关记忆 (Divide-and-Remember: Recursive Action-Relevant Memory for Long-Horizon VLA Policies) 叙事流：用机器人速度场做流匹配的 VLA 模型 (NarrativeFlow: Flow-Based Vision-Language-Action Model Using Robot Velocity Fields) 评估 VLA 的「记忆」：MIKASA-Robo-VLA 长时序操作基准 (MIKASA-Robo-VLA: Benchmarking Memory in VLA Models for Long-Horizon Manipulation) 成功率就够了么？输入扰动如何改变 VLA 在桌面操作中的行为 (Is Success All You Need? Investigating the Impact of Input Perturbations on VLA Behaviour in Tabletop Manipulation Tasks) 用 EMG 与视觉标注做连续任务条件化：VLA 可以不止靠语言 (Continuous Conditioning of VLAs with Augmenting EMG and Visual Task Descriptors) 以運動為中心的未來動力學：MotionWeave 給 VLA 的世界模型減負 (MotionWeave: Learning Motion-Centered Future Dynamics for Vision-Language-Action Policies) 用运行时反馈自进化失败银行：让 VLA 安全约束"学进"策略里 (Learning from Runtime Feedback through Failure-Bank Self-Evolution for Vision-Language-Action Models) 记住你做过什么：动作历史记忆 + 双专家去噪的长时序 VLA 策略 (Remember What You Did: Action-History Memory with Dual-Expert Denoising for Long-Horizon VLA Policies) 驯服执行误差下的 VLA：部署期自补偿与压力测试 (Taming VLAs under Robot Execution Errors: Self-Compensation and Stress Testing) 超越 Token 重要性：为高效 VLA 推理保留空间骨架 (Beyond Token Importance: Preserving Spatial Scaffolds for Efficient Vision-Language-Action Inference) ActionGround：冻结 VLA 策略的免训练运行时精炼 (ActionGround: Training-Free Runtime Refinement of Frozen VLA Policies) 把末端执行器「隐身」：Cloak 用几何遮挡实现零样本跨本体操作 (Cloak: Zero-Shot Cross-Embodiment Manipulation by Masking the End-Effector from the VLA) 神经动作编解码器：用音频 Codec 重做 VLA 的动作分词 (NAC: Neural Action Codec for Vision-Language-Action Models) 动作表示的「方向–尺度解耦」（Direction-Scale Decomposition, DSD）：重新审视离散 token VLA 该 tokenize 什么 RotVLA：旋轉潛動作——讓 Latent Action 住進 SO(n) (RotVLA: Rotational Latent Action for Vision-Language-Action Model) 自適應 VLA：讓部署後的策略用「自己的失敗軌跡」逐步補償硬體漂移 (Self-Adaptive VLA for Robust Robot Deployment) StageCraft：執行期干擾物與阻塞失效的緩解 (StageCraft: Execution Aware Mitigation of Distractor and Obstruction Failures in VLA Models) 单鱼眼解耦覆盖度与锐度：Fisheye-VLA (Fisheye-VLA: Decoupling Coverage and Acuity for Manipulation with a Single Fisheye Camera) RoboTwin-Phys：WAM 与 VLA 真的理解物理世界吗？ (RoboTwin-Phys: Do WAMs and VLAs Understand the Physical World?) 重建误差不等于控制性能：自回归 VLA 的动作 Tokenization 该看什么 (Beyond Reconstruction Error: Analytical and Data-Driven Action Tokenization for Autoregressive Vision-Language-Action Models) 给冻结 VLA 装一圈「安全带」：SafeLoop 的风险预测 + 回滚恢复 (SafeLoop: Risk-Aware Rollback for Vision-Language-Action Manipulation) 异步回放锚定的在线 VLA 后训练 (Stable and Efficient Real-World Online VLA Post-Training via Asynchronous Replay-Anchored Policy Improvement) 闭环崩塌解剖学：一次压缩 VLA 策略的因果个案研究 (Anatomy of a Closed-Loop Collapse: A Causal Case Study of a Compressed VLA Policy) 紧凑 VLA 的持久循环记忆：SmoLSTM (SmoLSTM: A Compact Vision-Language-Action Model with Recurrent Memory that Persists) CommitFlow：长程 VLA 执行的语义承诺验证与局部纠偏 (CommitFlow: Semantic Commitment Verification and Local Correction for Long-Horizon Robot Manipulation VLA Execution) 把人类在环后训练搬下机器人：HIL-UMI 的机器人免部署 VLA 后训练 (HIL-UMI: Bringing Human-in-the-Loop Post-Training of Vision-Language-Action Models to Universal Manipulation Interface) 用去噪轨迹的几何自适应切分动作块 (GeoAAC: Geometry-Based Adaptive Action Chunking from Denoising Trajectories in VLA Policies) ActionPiece：重新思考自回归 VLA 的动作分词 (ActionPiece: Rethinking Action Tokenization for Autoregressive Vision-Language-Action Models) 解耦视觉、语言与动作：高效多任务机器人策略 (Decoupling Vision, Language, and Action for Efficient Multi-Task Robot Policies) 云边交错：VLA-ULAP 用 7.4M 轻量预测器砍掉近半云端 VLA 调用 (VLA-ULAP: Interleaving Cloud VLA Calls with Ultra-Lightweight Local Action Prediction at the Edge) 把幾何接進動作專家：sensVLA 的空間接地 VLA 架構 (sensVLA: Spatially-Grounded Vision-Language-Action Model for Autonomous Wheel Loader) FluxVLA Engine：面向具身智能的一站式 VLA 工程平台 (FluxVLA Engine: A One-Stop VLA Engineering Platform for Embodied Intelligence) 神经符号 VLA：把操作拆成「原始動作計劃」再餵給連續控制器 (NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models)
🔬 Frontier Research · 25 篇篇
分階段操作：以行為相位解耦實現類人機器人操作 (Move-Then-Operate: Behavioral Phasing for Human-Like Robotic Manipulation) 更少 Token，更好动作：PyRUA-Lean 用代码执行接口让机器人智能体成功率 +14%、输入 Token −65% (Fewer Tokens, Better Action: GPT-6 Astra Robot Agents with 14% Higher Success Rate but 65% Fewer Tokens) Recova：代理引導的自主機器人操作失敗恢復 (Recova: Agent-Guided Failure Recovery for Autonomous Robotic Manipulation) SlotVLA：用物体-关系槽位替代稠密 Token 的机器人操作表征 (SlotVLA: Towards Modeling of Object-Relation Representations in Robotic Manipulation) 3DROID：可渲染 3D Gaussian 機器人數據集與逐場景可靠性度量 (3DROID: A Renderable 3D Gaussian Dataset with Measured Per-Scene Reliability) LiteReality-Agent：把 3D 重建变成一个「写代码」的 Agentic 系统 (LiteReality-Agent: An Agentic System for Interactable 3D Indoor Scene Reconstruction) DuoMind：以語義通訊實現分散式多機器人協調 (DuoMind: Enabling Distributed Multi-Robot Coordination with Semantic Communication) PhysicsLENS：診斷視頻生成模型的「物理屬性盲區」 (PhysicsLENS: Diagnosing Physical Property Blindness in Video Generation Models) 生物物理線蟲神經環路作為視覺魯棒機器人操作的「任務無關動力學核心」 (A Biophysically Detailed C. elegans Circuit as a Task-Agnostic Dynamical Core for Visually Robust Robot Manipulation) 自我中心人类演示直迁人形全身协调技能 (EgoHumanoid-V2: Human-to-Humanoid Transfer of Coordinated Whole-Body Skills for Loco-Manipulation) 从视频到形式化规范：Video2STL 如何把「看会的动作」变成可训练的信号 (Video2STL: Grounding VLM-Generated Temporal Specifications for Robot Learning) DeltaWAM：以 Delta Token 為單位的變更中心世界動作模型 (DeltaWAM: Change-Centric Visual Foresight via Delta Tokens for an Efficient World-Action Model) 以 3D Grounding 作为机器人基座：Grounded Action Model (Grounded Action Model: 3D Grounding as a Foundation for Robotics) 在动作流形上用黎曼 MeanFlow 加速视觉运动策略 (Faster Visuomotor Policy Learning on Action Manifolds via Riemannian MeanFlow) EgoSpeedUp：把人类操作节奏迁移给机器人策略 (EgoSpeedUp: Transferring Human Manipulation Tempo to Robot Policies) 掌中世界：野外人类操作数据的规模化生态 (World In Your Hands: A Large-Scale and Open-Source Ecosystem for Learning Human-Centric Manipulation in the Wild) LIFD：以锚定扩散构建 3D 感知场景记忆的操作策略 (LIFD: Anchored Diffusion for 3D-Aware Scene Memory in Robotic Manipulation) Workspace Models：用顯著性監督把 VLM 記憶「攤銷」進一個輕量 token (Workspace Models: Lightweight Robotic Memory via Saliency-Driven Supervision) 「你的機器人是在謊言上訓練的」：機械臂操作中的碰撞網格投毒攻擊 (Collision Mesh Poisoning Attacks on Robotic Manipulation) 障礙感知外殼：讓免訓練編碼智能體安全操作機械臂 (Coding Agents with an Obstacle-Aware Harness for Safe Robot Manipulation) AntiGrounding：把「可执行轨迹」当作视觉提示，让 VLM 选动作 (AntiGrounding: Executable Robot Trajectories as Visual Prompts for VLM-Guided Manipulation) DetAug：面向零样本避障的「障碍盲」轨迹增强 (Obstacle-Blind Trajectory Augmentation for Zero-shot Obstacle Avoidance) HumanEgo：用几分钟人类第一视角视频零样本学出机器人操作策略 (HumanEgo: Zero-Shot Robot Learning from Minutes of Human Egocentric Videos) 統一記憶-預測-動作：UniMPA 用「動作接地轉移 (Action-Grounded Transition)」縫合 VLA 的可執行性缺口 (UniMPA: A Unified Memory-Prediction-Action Model via Action-Grounded Transition Modeling) 用「可实现转移」打通预测、记忆与动作：UniMPA 深度拆解 (UniMPA: A Unified Memory-Prediction-Action Model via Action-Grounded Transition Modeling)
🌍 World Models · 14 篇篇
统一世界-动作模型：把「懂物理」和「会推理」焊进同一条 Transformer (UniWAM: Unified World-Action Model) SkeleWAM：骨架世界-動作建模 — 用稀疏 3D 骨架做高效機器人操作 (Skeleton World-Action Modeling for Efficient Robotic Manipulation) Magic-W0：把世界模型重构为「决策导向」的三段式结构（Magic-W0: A Structured World–Action Foundation Model for Physical Intelligence） 潛在世界模型的規劃極限 (The Planning Limits of Latent World Models) Q-WAM：面向世界动作模型的 4-bit 量化与动作子空间保护 (Q-WAM: 4-Bit Quantization of World Action Models with Action-Subspace Protection) 保留未来，丢掉推演：RIFT 让世界动作模型一步生成未来缓存 (Keep the Future, Drop the Rollout: RIFT for World Action Models) 在表征空间里直接规划：Representation World Model (Learning States, Transition and Executable Plans in Representation) 世界动作智能体：让 VLM 在「世界动作预演」中直接驾驶机器人 (World Action Agent: Harnessing VLMs for Robot Manipulation via World Action Rehearsal) 世界模型真的让机器人更强吗？——预测式具身智能评测基准综述 (Do World Models Make Better Robots? A Survey of Evaluation Benchmarks for Predictive Embodied Intelligence) 滚动世界-动作模型：把联合去噪摊到重规划周期 (Rolling-WAM: World Action Models with Rolling Imagination) 推理期仿真在环精修：让真实布料操作可以「现场纠错」 (Enabling Robust Cloth Manipulation via Inference-Time Simulator-in-the-Loop Refinement) 把交互动力学蒸馏进单步去噪：DIDO 让世界行动模型跑得更快 (DIDO: Distilling Interaction-Centric Dynamics into One-Step Denoising for World Action Models) 免訓練讀取世界模型內部態做失敗檢測 (FARM: Reading Failure Signals from the Internal Predictive States of a Frozen Robotic World Model) 冻结世界模型的内部预测状态里，已经写好了失败信号 (FARM: Reading Failure Signals from the Internal Predictive States of a Frozen Robotic World Model)
🌊 Diffusion & Flow · 9 篇篇
解耦式早退：為 Flow-Matching VLA 做任務相關的算力分配 (Decoupled Early Exits for Task-Dependent Compute Allocation in Flow-Matching VLAs) TraceFlow：用成功/失败轨迹引导冻结的流匹配机器人策略 (TraceFlow: Guiding Frozen Flow-Matching Robot Policies with Success and Failure Traces) Freeze, Share, Shrink：冻结、共享、压缩——重思扩散策略的动作骨干 (Freeze, Share, Shrink: Rethinking the Action Backbone in Diffusion Policies) Lie 群 B 样条扩散策略：让扩散策略的轨迹真正平滑 (LieSpline-DP: Lie-Group B-Spline Diffusion Policy for Smooth Robot Manipulation) AdaVLA：免训练自适应步数流匹配，让 VLA 推理提速 2 倍 (AdaVLA: Adaptive Step Flow Matching for Training-free Acceleration of Vision-Language-Action Models) ActSafeGuard：把硬约束注入流匹配策略的可微「射线缩放」安全层 (ActSafeGuard: Differentiable and Training-Aligned Constraint Enforcement for Flow-Matching Policies) IMLE-VLA：用单步 cIMLE 生成器替换迭代式动作头 (IMLE-VLA: Fast Single-Step Action Generation for Vision-Language-Action Policies) LTLf 约束引导的多智能体扩散策略 (LTLDiff: Finite Linear Temporal Logic-Guided Data Generation and Diffusion Policies for Multi-agent Robotic Manipulation) IMLE-VLA：用条件隐式最大似然估计实现单步动作生成 (IMLE-VLA: Fast Single-Step Action Generation for Vision-Language-Action Policies)
🤚 Tactile Perception · 9 篇篇
从高带宽单点传感学习触觉感知 (Learning tactile perception from high-bandwidth single-point sensing) 用触觉训练、用视觉部署：ZeroTouch 的「无触觉」接触力估计 (ZeroTouch: Tactile-Supervised Visual Contact Estimation for Contact-Rich Manipulation) 势场即动作：接触密集操作中 RL 的策略—运动解耦 (Potential-Field Action Representation for Reinforcement Learning in Contact-Rich Manipulation) 敏捷触觉世界动作模型：把触觉世界模型压进 ~10ms 推理 (Agile-WAM: An Agile Tactile World Action Model for Contact-Rich Robot Control) 人類觸覺澆灌機器人世界模型 (DexTouch-WM: Learning Action-Conditioned Tactile World Models from Human Touch for Dexterous Robot Manipulation) TAO-Force：融合力觉感知与快慢双环控制的接触丰富操作 (TAO-Force: Unifying Force-Aware Perception and Fast–Slow Control for Contact-Rich Manipulation) 接触丰富操作的 I/O 设计：命令-执行失配作为隐式力线索 (Mind the Gap: Rethinking I/O Design for Contact-Rich Visuomotor Policy Learning) Bench2Dex：跨 12 种灵巧手的双臂视觉-触觉操作基准 (Bench2Dex: Benchmarking Visuo-Tactile Bimanual Dexterous Manipulation Across Dexterous Hands) STAR：視-觸-語言-動作模型中的稀疏觸覺表徵學習（面向靈巧手操作）
🔧 Deployment & Hardware · 8 篇篇
MobiAgent：长时程移动操作的双环递归策略自改进 (MobiAgent: Dual-Loop Recursive Policy Self-Improvement for Long-Horizon Mobile Manipulation) DITTO-X：面向灵巧操作的正向与反向遥操作及人类干预 (DITTO-X: Forward and Reverse Teleoperation for Dexterous Manipulation and Human Intervention) FlashDexRetarget：多動作重定向加速靈巧操作資料生成 (FlashDexRetarget: Accelerating Dexterous Manipulation Data Generation through Multi-Motion Retargeting) VisForce：以視覺接地當前與期望力，實現目標條件式靈巧操作 (VisForce: Visual Grounding of Current and Desired Forces for Goal-Conditioned Dexterous Manipulation) MAVP：地圖感知視運動策略 — 讓移動操作機器人「知道自己站在哪」 (MAVP: Map-Aware Visuomotor Policies for Mobile Manipulation) 免费的柔顺：用双边遥操作学习可辨识阻抗 (Compliance for Free: Learning Identifiable Impedance via Bilateral Teleoperation) 高自由度灵巧手的 VLA 后训练四段式流程 (Towards High-DoF Dexterous Manipulation through VLA Post-Training) AnyViewDex：純 RGB 觀測下的視角不變靈巧操作 (AnyViewDex: View-Invariant Dexterous Manipulation from RGB Observations)
🧠 Planning & Reasoning · 6 篇篇
把前沿推理蒸馏进本地执行：自主演示生成 + 稠密语言监督 (Bridging Frontier Reasoning and Robot Execution: From Autonomous Demonstration Generation to Dense Language Supervision) 用「數位孿生」錨定真實影片預測：Real-Sim-Real 世界模型 (VPTwin: Real-Sim-Real Video Prediction for Robotic Manipulation Planning) CAPEX：用执行经验自适应推理，把多模态基础模型的行为蒸馏成可部署机器人策略 (CAPEX: Efficiently Distilling Foundation Model Behavior into Deployable Robot Policies through Experience-Adaptive Reasoning) 双路径动态操作：DualManip 用几何自适应绕开 VLM 延迟瓶颈 (DualManip: Agentic Dynamic Manipulation via Dual-Path Semantic Reasoning and Geometric Adaptation) RACaP：把「写代码」搬到演化期，运行时只调冻结的 Policy API (RACaP: Agentic Reasoning, Acting, and Coding as Policies for Evolvable Robot Learning) MotionForge：动态物体长时程操作的域偏移评测基准与数据生成管线 (MotionForge: A Data Generation Pipeline and Large-Scale Benchmark for Long-Horizon Manipulation of Dynamic Objects with Domain Shifts)
🎮 Reinforcement Learning · 5 篇篇
學習「何時」與「由哪個」RL 專家接管 VLA 策略 (RouteRLT: Learning When and Which RL Specialist Should Control a Vision-Language-Action Policy) ForceRFT：以力引导残差强化学习精修 VLA 动作 (ForceRFT: Refining VLA Actions through Force-Guided Residual Reinforcement Learning) 並行仿真中的視覺語言獎勵學習規模化 (Scaling Vision-Language Reward Learning for Robot Manipulation in Parallel Simulation) 复用 VLA 表征做内在奖励 (Intrinsic Robot Rewarding: Reusing VLA Representations for Autonomous Evaluation and Policy Improvement) 从噪声引导到双隐空间：给冻结生成式策略加一条「内部表征」控制通道 (Beyond Noise Steering: Dual-Latent Space Reinforcement Learning for Generative Robot Policy)
🏗️ Foundation & Training · 4 篇篇
GDLAM：分组解耦的隐动作模型，为具身预训练提供"可干预"的表征 (GDLAM: Group-Disentangled Latent Action Model for Highly Disentangled Embodied Pretraining) 几何感知潜在动作建模：让跨具身 VLA 预训练"看见手指" (GALA: Geometry-Aware Latent Action Modeling for Vision-Language-Action Model Pretraining across Embodiments) StarVLA-α：给 VLA 系统做减法 (StarVLA-α: Reducing Complexity in Vision-Language-Action Systems) ParticleSplat：自监督物体中心潜在粒子泼溅 (ParticleSplat: Self-supervised Object-centric Latent Particle Splatting)
👁️ Perception & 3D · 1 篇篇
🏆 SOTA 排行SOTA 排行
Evo-SOTA 完整榜Evo-SOTA 完整榜 →
30
CALVIN ABCD-D 飽和饱和 avg_len
| # | Model | Score | vs Prev | Date | Paper |
|---|---|---|---|---|---|
| 1 | Xiaomi-Robotics-0 | 4.8 | Flower VLA +0.13 | 2026-10-23 | arxiv → |
| 2 | Xiaomi-Robotics-0 | 4.8 | Flower VLA +0.13 | 2026-10-16 | arxiv → |
| 3 | MMaDA-VLA | 4.78 | Xiaomi-Robotics-0 +0.03 | 2026-10-23 | arxiv → |
| 4 | MMaDA-VLA | 4.78 | Xiaomi-Robotics-0 +0.03 | 2026-10-16 | arxiv → |
| 5 | EDAR | 4.75 | NIAF +0.09 | 2026-10-23 | arxiv → |
| 6 | EDAR | 4.75 | NIAF +0.09 | 2026-10-16 | arxiv → |
| 7 | AVA-VLA | 4.65 | NIAF +0.18 | 2026-10-23 | arxiv → |
| 8 | AVA-VLA | 4.65 | NIAF +0.18 | 2026-10-16 | arxiv → |
| 9 | NS-VLA | 4.56 | AtomicVLA +0.29 | 2026-10-23 | arxiv → |
| 10 | NS-VLA | 4.56 | AtomicVLA +0.29 | 2026-10-16 | arxiv → |
| 11 | HiMoE-VLA | 4.49 | Flower VLA +0.14 | 2026-10-23 | arxiv → |
| 12 | HiMoE-VLA | 4.49 | Flower VLA +0.14 | 2026-10-16 | arxiv → |
| 13 | MCIL | 1.82 | 2026-10-23 | arxiv → | |
| 14 | MCIL | 1.82 | 2026-10-16 | arxiv → |