Learning Diffusion Policy from Primitive Skills for Robot Manipulation
这是一份面向零基础读者的结构化研究笔记。本文只记录公开论文文本、arXiv 元数据和公开摘要能支持的结论;本站没有复现 CALVIN、LIBERO 或真实机器人实验,因此不会把论文报告的成功率写成本站 E4 结果。
一句话讲什么(TL;DR)
Learning Diffusion Policy from Primitive Skills for Robot Manipulation 提出 Skill-conditioned Diffusion Policy,简称 SDP。它认为很多 diffusion policy 直接从高层语言指令生成短期动作,会出现粒度错位:语言说“把柠檬放进锅里”,但机器人每一刻真正要做的是 move up、move down、open gripper、close gripper 这类短程基础动作。
SDP 抽象出八个 reusable primitive skills,并用 VLM 从视觉观察和语言指令中提取离散表示,再通过 lightweight router 为每个状态选择一个 primitive skill。之后,skill-conditioned diffusion policy 根据选中的 skill 生成 skill-aligned actions。论文报告 SDP 在 CALVIN、LIBERO 和真实机器人任务上超过 SOTA baseline,例如 CALVIN ABC→D 五连任务成功率 76.9%,LIBERO 平均成功率 96.9%。
如果只记一个直觉:SDP 给机器人加了一层“动作短语”。人说的是完整任务,低层 policy 输出的是关节动作,中间用“打开夹爪、向上移动、旋转”这类 primitive skills 对齐。
所以这一节是想说:SDP 用可解释 primitive skills 缓解语言指令和低层动作之间的粒度错位。
这是个什么场景
机器人操作任务常常是长程组合任务。比如 “Pick up the lemon and put it into the yellow pan” 看起来是一句话,但实际包含接近、下移、张开夹爪、闭合夹爪、抬起、移动、放下等多个短动作。
普通 diffusion policy 可以生成动作序列,但如果它只接收全局高层指令,就要自己隐式学会这些短动作阶段。模型可能知道最终目标,却在某一刻不知道该闭夹爪还是移动,导致 action generation misalignment。
SDP 的场景就是把复杂任务拆成更短、更可解释、更可复用的 primitive skills。它不是让人手写完整任务程序,而是让模型动态选择当前状态该用哪个技能。
高层任务和低层动作之间的粒度差
High-level instruction:
"Pick up the lemon and put it into the yellow pan"
│
▼
Primitive skill sequence:
move down -> open gripper -> close gripper -> move up -> translate -> move down
│
▼
Low-level actions:
continuous robot control signals
所以这一节是想说:长程操作需要中间层,否则语言和动作粒度差太大。

之前的人怎么做的,为什么不够好
原始 diffusion policy 从视觉观察和动作示范中学习动作分布,对单任务或短程控制很强。但在多任务语言条件操作中,模型需要把高层语言直接映射到低层动作,这很难。
一些方法加入语言 encoder,把 instruction embedding 和 noisy action sequence 一起输入 diffusion policy。这样能处理更多任务,但语言 embedding 往往太粗,不一定告诉模型当前时刻应该做哪个短动作。
还有一些 VLA 或 multitask policy 学隐式技能表示。隐式表示可能有效,但很难解释和调试。机器人失败时,我们不知道它是“没理解任务”,还是“选错了当前技能”,还是“低层动作生成错了”。
SDP 的判断是:primitive skills 应该显式出现。它们既能被人理解,又能作为 action generation 的条件,让 diffusion policy 更稳定地产生 skill-consistent behavior。
所以这一节是想说:旧方法把技能藏在模型里,SDP 想把技能显式拿出来。
这篇论文的新想法
第一,新想法是八个 reusable primitive skills。论文列出 roll、yaw、open the gripper、move up、translate、close the gripper、move down、rotate。这些是跨任务共享的短程 manipulations。
第二,新想法是 lightweight router。模型从视觉观察和高层语言指令中提取 representation,然后 router 选择当前状态最合适的 primitive skill。
第三,新想法是 skill-conditioned diffusion policy。选中的 skill 不只是标签,而是参与 action generation。论文设计 skill-dependent FFN layer,用 skill embedding 动态影响 diffusion policy。
第四,新想法是可解释性。每个状态分配一个 primitive skill,研究者可以检查机器人当前“以为自己在做什么”。这比完全端到端的隐式控制更容易分析。
SDP 框架
Visual observations + language instruction
│
▼
VLM / representation encoder
│
▼
Lightweight router selects primitive skill
│
▼
Skill-conditioned diffusion policy
│
▼
Skill-aligned continuous actions
所以这一节是想说:SDP 把 skill selection 和 diffusion action generation 组合成一个层次化但可训练的系统。
它分几步做的(方法)
第 1 步:定义 primitive skill set
输入是多种 robot manipulation tasks。作者观察这些任务可以拆成基础短动作,于是定义八个 primitive skills:roll、yaw、open the gripper、close the gripper、move up、move down、translate、rotate。
处理过程是把高层任务的连续控制过程看成 primitive skill 序列。输出是一个共享 skill set。这个 set 不绑定某个具体任务,而是跨任务复用。
这一步的价值是提供中间语言。人能理解这些 skill,模型也可以用它们作为条件,降低直接从 task instruction 到 action 的难度。
第 2 步:用 compositional prompt ensemble 构造 skill embeddings
论文使用统一模板,例如 “the robot arm is going to {skill}”。每个 primitive skill 都放进这个模板,经过 CLIP text encoder 和 MLP,得到 skill prompt embeddings。
输入是八个 skill 文本。处理是 text encoding 和 MLP projection。输出是八个 skill embeddings,供 router 和 diffusion policy 使用。
这一步让 skill 不只是字符串,而是进入模型计算图的向量表示。
第 3 步:从视觉和语言中提取状态表示
输入包括 static camera、wrist camera 等视觉观察,以及高层语言指令。模型使用 vision-language model 提取 discrete representations 或 multimodal representations。
处理过程是编码视觉和语言上下文,得到当前状态的综合表示。输出是 router 可以判断的 state representation。
这一步解决“当前该做什么”的信息来源。仅看语言不够,因为同一句任务在不同阶段要做不同 skill;仅看图像也不够,因为不知道目标。
第 4 步:router 为每个状态选择 primitive skill
Lightweight router network 输入 state representation,输出八个 primitive skills 的 importance scores。最终通过 top-1 selection 选择当前状态的 desired primitive skill。
处理过程是动态 skill assignment。输出是一个具体 skill,比如 close gripper 或 move up。
这一步是 SDP 的决策中间层。它把长程任务拆成一连串状态相关的 skill choices。
第 5 步:skill-dependent FFN 影响 diffusion policy
选中的 skill embedding 会进入 diffusion policy。论文引入 skill-dependent feed-forward network layer,类似 LoRA-like 动态参数化,让 diffusion policy 根据 skill 调整动作生成。
输入是 noisy actions、observation features、skill embedding。处理是 diffusion denoising,同时用 skill-conditioned layer 调整中间表示。输出是 continuous actions。
这一步让 skill 真正影响低层动作,而不只是附在输入里的标签。比如 open gripper 和 move down 对同一视觉状态应该产生完全不同动作。
第 6 步:在 CALVIN、LIBERO 和真实机器人上评估
CALVIN 测长程语言条件操作,尤其 ABC→D zero-shot generalization。LIBERO 包含 Spatial、Object、Goal、Long 四类 task suites。真实机器人实验则检验 sim-to-real 和视觉干扰下的表现。
论文报告 SDP 在 CALVIN ABC→D 五连任务成功率 76.9%,超过 MoDE 14.5%、UniVLA 20.4%;LIBERO 平均成功率 96.9%,超过 MDT 13.4%、UniVLA 4.4%,且是唯一在 LIBERO-Long 超过 90% 的方法。
这些结果说明 primitive skill 作为中间层能提升长程组合和多任务泛化,但所有数字仍是论文报告。
所以这一节是想说:SDP 的方法是“选当前 primitive skill,再让 diffusion policy 按 skill 生成动作”。

关键数字
| 数字 | 原文语境 | 这说明什么 |
|---|---|---|
| 8 | reusable primitive skills | 技能中间层规模 |
| 4 | SDP 生成动作使用的 denoising steps | 少于一些 diffusion baseline 的 10 steps |
| 76.9% | CALVIN ABC→D 五连任务成功率 | 论文报告 zero-shot generalization 强 |
| 14.5% | 相比 MoDE 的提升 | CALVIN ABC→D 设置 |
| 20.4% | 相比 UniVLA 的提升 | CALVIN ABC→D 设置 |
| 96.9% | LIBERO 四套件平均成功率 | 论文报告多任务表现 |
| 13.4% | 相比 MDT 的 LIBERO 平均提升 | baseline 对比 |
| 4.4% | 相比 UniVLA 的 LIBERO 平均提升 | baseline 对比 |
这些数字全部是论文报告,不是本站复现实验。后续人工核验应回到 Table 1、Table 2 和真实机器人图表逐项确认。
所以这一节是想说:SDP 的证据集中在 CALVIN/LIBERO 长程和多任务成功率。
实验结果说明了什么
实验说明 primitive skills 对长程任务有帮助。CALVIN 要连续完成多个 instruction,如果每一步都从高层语言直接生成动作,模型容易漂移;SDP 用 skill sequence 给动作生成提供更细粒度约束。
LIBERO 结果说明,skill-conditioned policy 不只适合一个 benchmark。Spatial、Object、Goal、Long 四类任务对空间关系、对象变化、目标变化和长程组合都有要求,SDP 的平均成功率高说明中间 skill 有泛化价值。
Ablation 结果也重要。论文报告 skill-dependent FFN、explicit skill abstraction、compositional prompt ensemble 都有贡献。也就是说,不只是“随便加个 skill label”就有效,skill 如何进入 diffusion policy 很关键。
真实机器人实验说明,primitive skill abstraction 对部署有实际意义。可解释技能使得人可以更容易观察机器人当前阶段是否合理。
更重要的是,SDP 把失败原因变得更可定位。端到端 policy 失败时,我们只知道动作错了,却很难知道模型内部阶段是否正确。引入 primitive skill 后,可以先看 router 是否选错技能,再看 diffusion policy 是否在正确技能下生成了错误轨迹。前者是阶段理解问题,后者是低层控制问题。
这种可分解性对长程任务尤其有价值。比如同样是“把物体放进容器”,失败可能发生在接近、抓取、抬起、移动、释放任一阶段。Primitive skill 序列能作为一种轻量任务进度条,让研究者更容易发现任务在哪个阶段偏离,也更容易针对性收集数据或调模型。
所以这一节是想说:实验支持“技能中间层”能缓解语言到动作的粒度错位。
术语表
- Primitive skill:短程、细粒度、可复用的基础操作,如 move up、close gripper。
- Skill-conditioned diffusion policy:以 skill 作为条件生成动作的 diffusion policy。
- Router network:根据当前视觉和语言状态选择 skill 的轻量网络。
- Skill-dependent FFN:根据 skill embedding 动态影响 diffusion policy 的前馈层。
- CALVIN:长程语言条件机器人操作 benchmark。
- LIBERO:测试知识迁移和多任务机器人操作的 benchmark。
- LoRA-like:类似低秩适配的参数调制思想。
所以这一节是想说:SDP 的关键词是 primitive skill、router 和 skill-conditioned denoising。
局限和边界
第一,八个 primitive skills 是否足够覆盖所有真实任务仍不确定。复杂接触、双臂协作、工具使用可能需要更丰富 skill set。
第二,router 选错 skill 会直接影响动作生成。如果当前应该 close gripper,却选成 translate,policy 会产生错动作。
第三,primitive skill 的可解释性不等于安全性。人能看懂 skill,不代表动作一定安全。
第四,CALVIN/LIBERO 是重要 benchmark,但真实环境中物体、光照、传感器和动力学差异更大。
第五,SDP 仍依赖示范数据和 diffusion policy 能力。没有足够数据的技能,不能靠 skill 名字凭空学会。
第六,skill assignment 也可能受视觉遮挡影响。如果关键物体被挡住,router 可能无法判断当前阶段,低层 policy 即使能力足够也会接到错误条件。
因此真实系统通常还需要失败检测、重新观察和人工接管,而不是只依赖一次 skill 选择。
所以这一节是想说:SDP 给了更好的中间层,但 skill set、router 和真实部署仍要验证。
和其他论文的关系
和 disco-diffusion-policy 相比,SDP 用 primitive skills 作为中间层,DISCO 用 VLM-generated keyframes 作为中间层。
和 time-unified-diffusion-policy 相比,SDP 关注 task/action granularity,TUDP 关注 denoising efficiency。
和 trace-focused-diffusion-policy 相比,SDP 用当前 skill disambiguate action,TF-DP 用历史 trace disambiguate long-horizon stage。
和原始 diffusion-policy 相比,SDP 把高层 instruction 到动作的映射拆成更可解释的 skill-conditioned generation。
所以这一节是想说:SDP 是 Batch 7 里“技能层次化 diffusion policy”的代表。
和本导读的关系
本站学习 diffusion policy 时,容易把 policy 看成一个黑盒:输入图像和语言,输出动作。SDP 提醒我们,中间表示很重要。可解释 primitive skills 能降低黑盒程度,也能帮助 debugging。
它适合和 imitation learning、VLA、skill learning、hierarchical policy 一起读。读者可以思考:未来通用机器人是不是需要一套可学习、可组合、可解释的 skill vocabulary?
所以这一节是想说:SDP 让 diffusion policy 更接近可解释的长程操作系统。
思考题
- 为什么高层语言指令直接生成短期动作容易粒度错位?
- 八个 primitive skills 为什么能跨任务复用?
- Router network 需要同时看视觉和语言,原因是什么?
- Skill-dependent FFN 比简单拼接 skill embedding 可能强在哪里?
- 如果一个任务需要“擦拭”或“扭瓶盖”,现有 skill set 是否足够?
FAQ
Q:SDP 是不是手写规则系统? A:不是。Primitive skills 是显式中间层,但 router 和 skill-conditioned diffusion policy 都是学习得到的。
Q:八个 skill 是机器人最终动作吗? A:不是。它们是短程语义动作类别,最终仍由 diffusion policy 生成连续控制信号。
Q:96.9% 是本站复现的吗? A:不是。它是论文在 LIBERO 上报告的平均成功率。
Q:可解释 skill 是否保证机器人安全? A:不保证。它只提高可观察性和结构性,安全仍需要约束、监控和验证。
进一步读什么
diffusion-policy:理解动作扩散基础。- CALVIN / LIBERO:理解长程语言条件操作 benchmark。
- MoDE / MDT:理解 diffusion policy 的多任务和 transformer baseline。
trace-focused-diffusion-policy:比较 skill condition 和 trace condition。
精读补充:primitive skill 为什么既是能力层,也是调试层
SDP 的 primitive skill 不只是为了提高成功率,也是在给机器人行为加一个可检查的中间状态。端到端 policy 失败时,我们通常只能看到动作错了,却不知道模型内部为什么错。引入 skill 后,失败可以拆成两类:router 是否选错 skill,或者 diffusion policy 是否在正确 skill 下生成了错误动作。
这种可分解性对长程任务很重要。比如机器人把柠檬放进锅里失败,可能是“应该 close gripper 时还在 translate”,也可能是“close gripper 已经选对,但夹爪轨迹太偏”。前者是阶段理解问题,后者是低层控制问题。Primitive skill 让这两个问题更容易分开诊断。
不过,primitive skill 也会带来抽象边界。八个技能听起来通用,但真实世界的动作远比八类丰富。擦拭、插入、旋拧、拉链、柔性物体操作都可能需要更细或不同的 skill vocabulary。如果 skill set 太粗,router 即使选对也无法给 policy 足够指导;如果 skill set 太细,router 学习会变难。
因此,SDP 的更大启发不是“八个技能就是最终答案”,而是“语言到动作之间应该有可学习、可解释、可组合的中间层”。未来系统可能自动发现 skill,也可能允许人类定义一部分技能,再让模型学习组合。
所以这一节是想说:primitive skills 同时提升动作条件、长程分解和失败可诊断性。
后续核验清单
如果之后要把本文从 UNVERIFIED 提升到人工核验状态,应逐项核对:八个 primitive skills 的原文名称;CPE 模板;router top-1 selection;skill-dependent FFN 设计;CALVIN ABC→D 76.9%、MoDE +14.5%、UniVLA +20.4%;LIBERO 96.9%、MDT +13.4%、UniVLA +4.4%;真实机器人任务设置。
原文信息
- arXiv: 2601.01948
- PDF: https://arxiv.org/pdf/2601.01948
@article{gu2026primitiveskills,
title = {Learning Diffusion Policy from Primitive Skills for Robot Manipulation},
author = {Gu, Zhihao and Yang, Ming and Zou, Difan and Xu, Dong},
journal = {arXiv preprint arXiv:2601.01948},
year = {2026}
}
◼
引用本笔记 / Cite this note
@online{eai_primitive_skill_diffusion_policy_2026,
title = {(readable note) Learning Diffusion Policy from Primitive Skills for Robot Manipulation},
author = {Xun, Jason},
year = {2026},
note = {Note on a 2026 paper},
howpublished = {\url{https://estelledc.github.io/embodied-ai-reading-station/papers/primitive-skill-diffusion-policy/}},
organization = {Embodied AI: Zero to One}
}
All 202 papers (full index)
- 1. LLaVA: Visual Instruction Tuning
- 2. 3DShape2VecSet: 3D Shape Representation for Diffusion Models
- 3. SayCan: Do As I Can, Not As I Say
- 4. OpenVLA: An Open-Source Vision-Language-Action Model
- 5. VLAS: VLA Model With Speech Instructions
- 6. MLA: Multisensory Language-Action Model
- 7. Cosmos Policy: Fine-Tuning Video Models for Visuomotor Control
- 8. CartoRadar: RF-Based 3D SLAM Rivaling Vision Approaches
- 9. mmCLIP: Boosting mmWave-based Zero-shot HAR via Signal-Text Alignment
- 10. mmNorm: Non-Line-of-Sight 3D Object Reconstruction via mmWave Surface Normal Estimation
- 11. Proactive Hearing Assistants that Isolate Egocentric Conversations
- 12. NeuralAids: Wireless Hearables With Programmable Speech AI Accelerators
- 13. Creating speech zones with self-distributing acoustic swarms
- 14. Conv-TasNet: Surpassing Ideal Time-Frequency Magnitude Masking for Speech Separation
- 15. SoundStream: An End-to-End Neural Audio Codec
- 16. AudioLM
- 17. Conformer
- 18. Dual-path RNN
- 19. EnCodec
- 20. Meta-StyleSpeech
- 21. MusicLM
- 22. Robust Speech Recognition via Large-Scale Weak Supervision
- 23. SeamlessM4T
- 24. Stable Audio
- 25. Universal Source Separation with Weakly Labelled Data
- 26. Meta-World: A Benchmark and Evaluation for Multi-Task and Meta Reinforcement Learning
- 27. RLBench: The Robot Learning Benchmark & Learning Environment
- 28. robosuite: A Modular Simulation Framework and Benchmark for Robot Learning
- 30. CALVIN
- 31. LIBERO
- 32. RH20T
- 33. What Matters in Learning from Offline Human Demonstrations for Robot Manipulation
- 34. DROID
- 35. Open X-Embodiment
- 36. RoboCasa
- 37. SimplerEnv
- 38. Diffusion Policy: Visuomotor Policy Learning via Action Diffusion
- 39. 3D Diffusion Policy: Generalizable Visuomotor Policy Learning via Simple 3D Representations
- 40. Consistency Policy: Accelerated Visuomotor Policies via Consistency Distillation
- 41. EquiBot: SIM(3)-Equivariant Diffusion Policy
- 42. DiT-Policy
- 43. Diffusion Policy Policy Optimization (DPPO)
- 44. Affordance-based Robot Manipulation with Flow Matching
- 45. FlowPolicy: 3D Flow-based Policy via Consistency Flow Matching
- 46. FAST: Efficient Action Tokenization for VLA
- 47. π₀: A Vision-Language-Action Flow Model for General Robot Control
- 48. pi_0.5: VLA with Open-World Generalization
- 49. A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning
- 50. Generative Adversarial Imitation Learning
- 51. Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (ACT/ALOHA)
- 52. AnyTeleop
- 53. Behavior Transformers: Cloning k Modes with One Stone
- 54. Implicit Behavioral Cloning
- 55. RoboCat
- 56. ALOHA 2
- 58. HumanPlus
- 59. Generalizable Humanoid Manipulation with 3D Diffusion Policies (iDP3)
- 60. Mobile ALOHA
- 61. SmolVLA
- 62. Universal Manipulation Interface
- 63. Behavior Generation with Latent Actions (VQ-BeT)
- 64. ImageBind: One Embedding Space To Bind Them All
- 65. Connecting Touch and Vision via Cross-Modal Prediction
- 66. AnyMAL: An Efficient and Scalable Any-Modality Augmented Language Model
- 67. AudioPaLM
- 68. FROMAGe: Grounding LLMs to Images
- 69. OneLLM
- 70. X-VLM: Multi-Grained Vision Language Pre-Training
- 71. Tactile Beyond Pixels (Sparsh-X)
- 72. Sparsh: Self-supervised Touch Representations
- 73. Tactile-VLA
- 74. TLA: Tactile-Language-Action
- 75. Code as Policies: Language Model Programs for Embodied Control
- 76. Inner Monologue: Embodied Reasoning through Planning with Language Models
- 77. LLM+P: Empowering LLMs with Optimal Planning
- 78. PaLM-E: An Embodied Multimodal Language Model
- 79. ProgPrompt
- 80. ChatGPT for Robotics
- 81. GenSim
- 82. RoboFlamingo
- 83. Tree-Planner
- 84. VoxPoser
- 85. See Through Smoke: Robust Indoor Mapping with Low-cost mmWave Radar
- 86. Can WiFi Estimate Person Pose?
- 87. 3DRIMR: 3D Reconstruction and Imaging via mmWave Radar based on Deep Learning
- 88. milliEgo: Single-chip mmWave Radar Aided Egomotion Estimation via Deep Sensor Fusion
- 89. High Resolution Point Clouds from mmWave Radar
- 90. RadarSLAM: Radar based Large-Scale SLAM in All Weathers
- 91. Through-Wall Pose Imaging in Real-Time with a Many-to-Many Encoder/Decoder Paradigm
- 92. RFMask: A Simple Baseline for Human Silhouette Segmentation with Radio Signals
- 93. RFPose-OT: RF-Based 3D Human Pose Estimation via Optimal Transport Theory
- 94. Argus: Multi-View Egocentric Human Mesh Reconstruction Based on Stripped-Down Wearable mmWave Add-on
- 95. Diffusion Model is a Good Pose Estimator from 3D RF-Vision
- 96. Enabling Visual Recognition at Radio Frequency (PanoRadar)
- 97. Wave-Former: Through-Occlusion 3D Reconstruction via Wireless Shape Completion
- 98. Habitat: A Platform for Embodied AI Research
- 99. Isaac Gym: High Performance GPU-Based Physics Simulation For Robot Learning
- 101. Habitat 2.0
- 102. ManiSkill
- 103. ProcTHOR
- 104. SAPIEN: A SimulAted Part-based Interactive ENvironment
- 105. BEHAVIOR-1K
- 106. BridgeData V2
- 106. Habitat 3.0
- 107. Isaac Lab
- 108. DexMV
- 108. MuJoCo Playground
- 109. DexCap
- 109. RT-1: Robotics Transformer for Real-World Control at Scale
- 110. 3D Diffusion Policy (DP3)
- 111. Octo: An Open-Source Generalist Robot Policy
- 112. RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control
- 113. RT-Trajectory: Robotic Task Generalization via Hindsight Trajectory Sketches
- 114. 3D-VLA
- 116. GR-2: Generative Video-Language-Action Model
- 117. DexVLA
- 117. OpenHelix
- 118. Cosmos World Foundation Model
- 118. OpenVLA-OFT
- 119. RDT-1B: Diffusion Foundation Model for Bimanual Manipulation
- 120. RoboMamba
- 121. SpatialVLA
- 122. TinyVLA
- 123. TraceVLA: Visual Trace Prompting
- 124. Learning Transferable Visual Models From Natural Language Supervision
- 125. Flamingo: a Visual Language Model for Few-Shot Learning
- 126. BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models
- 127. BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation
- 128. DeepSeek-VL: Towards Real-World Vision-Language Understanding
- 129. EVA-CLIP: Improved Training Techniques for CLIP at Scale
- 130. FILIP: Fine-grained Interactive Language-Image Pre-Training
- 131. Florence-2: Advancing a Unified Representation for a Variety of Vision Tasks
- 132. InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks
- 133. Improved Baselines with Visual Instruction Tuning
- 134. OBELICS
- 135. Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond
- 136. Sigmoid Loss for Language Image Pre-Training
- 137. What matters when building vision-language models?
- 138. Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling
- 139. The Llama 3 Herd of Models
- 140. LLaVA-NeXT-Interleave
- 141. LLaVA-OneVision: Easy Visual Task Transfer
- 142. Long-CLIP: Unlocking the Long-Text Capability of CLIP
- 143. Pixtral 12B
- 144. Dream to Control: Learning Behaviors by Latent Imagination
- 145. World Models
- 146. DayDreamer
- 147. Mastering Atari with Discrete World Models
- 148. Dreamer V3: Mastering Diverse Domains through World Models
- 149. Transformers are Sample-Efficient World Models
- 150. TWM: Transformer-based World Models
- 151. 1X World Model Challenge
- 153. GAIA-1
- 154. Genie: Generative Interactive Environments
- 155. Navigation World Models
- 156. UniSim
- 157. LeRobot: An Open-Source Library for End-to-End Robot Learning
- 158. CogACT: A Foundational Vision-Language-Action Model for Synergizing Cognition and Action in Robotic Manipulation
- 159. Universal Actions for Enhanced Embodied Foundation Models
- 160. LoHoVLA: A Unified Vision-Language-Action Model for Long-Horizon Embodied Tasks
- 161. AutoRT: Embodied Foundation Models for Large Scale Orchestration of Robotic Agents
- 162. EO-1: Interleaved Vision-Text-Action Pretraining for General Robot Control
- 163. Qwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot Embodiments
- 164. RealMirror: A Comprehensive, Open-Source Vision-Language-Action Platform for Embodied AI
- 165. LLaDA-VLA: Vision Language Diffusion Action Models
- 166. Discrete Diffusion VLA: Bringing Discrete Diffusion to Action Decoding in Vision-Language-Action Policies
- 167. Vlaser: Vision-Language-Action Model with Synergistic Embodied Reasoning
- 168. X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model
- 169. Embodiment Transfer Learning for Vision-Language-Action Models
- 170. HiMoE-VLA: Hierarchical Mixture-of-Experts for Generalist Vision-Language-Action Policies
- 171. Green-VLA: Staged Vision-Language-Action Model for Generalist Robots
- 172. AC^2-VLA: Action-Context-Aware Adaptive Computation in Vision-Language-Action Models for Efficient Robotic Manipulation
- 173. MoS-VLA: A Vision-Language-Action Model with One-Shot Skill Adaptation
- 174. Asynchronous Fast-Slow Vision-Language-Action Policies for Whole-Body Robotic Manipulation
- 175. VLA-Forget: Vision-Language-Action Unlearning for Embodied Foundation Models
- 176. Membership Inference Attacks on Vision-Language-Action Models
- 177. A Survey on Efficient Vision-Language-Action Models
- 178. Survey of Vision-Language-Action Models for Embodied Manipulation
- 179. Embodied AI with Foundation Models for Mobile Service Robots: A Systematic Review
- 180. Toward Embodied AGI: A Review of Embodied AI and the Road Ahead
- 181. RoboNeuron: A Middle-Layer Infrastructure for Agent-Driven Orchestration in Embodied AI
- 182. Embodied Navigation Foundation Model
- 183. MiMo-Embodied: X-Embodied Foundation Model Technical Report
- 184. Open-H-Embodiment: A Large-Scale Dataset for Enabling Foundation Models in Medical Robotics
- 185. AlanaVLM: A Multimodal Embodied AI Foundation Model for Egocentric Video Understanding
- 186. 3D Generation for Embodied AI and Robotic Simulation: A Survey
- 187. DISCO: Language-Guided Manipulation with Diffusion Policies and Constrained Inpainting
- 188. Time-Unified Diffusion Policy with Action Discrimination for Robotic Manipulation
- 189. Learning Diffusion Policy from Primitive Skills for Robot Manipulation
- 190. Trace-Focused Diffusion Policy for Multi-Modal Action Disambiguation in Long-Horizon Robotic Manipulation
- 191. Gaze2Act: Gaze-Conditioned Vision-Language-Action Policies for Interactive Robot Manipulation
- 192. LACY: A Vision-Language Model-based Language-Action Cycle for Self-Improving Robotic Manipulation
- 193. villa-X: Enhancing Latent Action Modeling in Vision-Language-Action Models
- 194. InstructVLA: Vision-Language-Action Instruction Tuning from Understanding to Manipulation
- 195. Discrete Policy: Learning Disentangled Action Space for Multi-Task Robotic Manipulation
- 196. Towards Generalizable Vision-Language Robotic Manipulation: A Benchmark and LLM-guided 3D Policy
- 197. A Survey of Language-Conditioned Robot Manipulation
- 198. SafeEmbodAI: a Safety Framework for Mobile Robots in Embodied AI Systems
- 199. The Essential Role of Causality in Foundation World Models for Embodied AI
- 200. A call for embodied AI
- 201. Toward General-Purpose Robots via Foundation Models: A Survey and Meta-Analysis
- 202. What Breaks Embodied AI Security: LLM Vulnerabilities, CPS Flaws, or Something Else?