Awesome Embodied Agents

A curated reading list on foundation-model-driven embodied agents: reasoning, planning, tool use, spatial memory, execution feedback, and self-improvement.

Focus: papers first published from September 2026 onward. Earlier papers are kept in a collapsed background-reading section. VLA and world-model work is included when it contributes to an agent system or provides a meaningful comparison.

Read the paper collection · 简体中文 · Suggest a paper · RSS

47 curated entries · 44 papers · 2 projects · 1 perspective

46 source-checked · 1 full-review · 0 AI drafts · 263 discovery candidates

• Recent papers
  • Reasoning & Planning · 3
  • Memory & Knowledge · 3
  • Skills & Code · 4
  • Self-improvement · 1
  • Benchmarks & Evaluation · 4
• Projects and systems
• Perspectives and essays
• Foundations and earlier work

Each paper appears once under its primary topic; secondary topics remain visible as tags. English titles link to original sources; Chinese notes and reading cards are available alongside them.

• BT-Agent: Deploying Onboard Robot Agents with Behavior Trees and Multi-Fidelity Evaluation
  2026-10-09 · Reasoning & Planning · Monitoring & Recovery · Benchmarks & Evaluation · Memory & Knowledge
  让机载小模型选择行为树技能，由本地检查及时中断异常执行，并用结构化反馈支持重试。
  

• NavGPT-3: Harnessing Context in a Hierarchical Navigation Runtime
  2026-10-07 · Reasoning & Planning · Monitoring & Recovery · Memory & Knowledge · VLA & World Models
  让推理、导航执行和监控拥有各自上下文，由运行时分配运动控制权，连接前沿模型与导航 VLA。
  

• RobotUse: Allocating Computation, Context, and Decisions
  2026-10-04 · Reasoning & Planning · Memory & Knowledge · Monitoring & Recovery · Skills & Code
  主 Agent 管目标，子 Agent 在图像中选点和姿态，几何后端执行；跨回合经验沉淀到可更新的操作手册。
  

• REMM: Revisable Embodied Memory for Manipulation
  2026-10-08 · Memory & Knowledge · Reasoning & Planning · Monitoring & Recovery · VLA & World Models
  给冻结 VLA 加一层可撤销的任务记忆：证据变了，依赖它的“已完成”状态也必须重审。
  

• Recursive Video In-Context Learning for Agentic Robot
  2026-10-05 · Memory & Knowledge · Reasoning & Planning · VLA & World Models · Monitoring & Recovery
  把一段任务示范变成可按需查询的层级视频记忆，让 Agent 在接触细节需要时再读取短片。
  

• RoboICL: Embodied In-Context Learning with GPT-6 Astra
  2026-09-28 · Memory & Knowledge · Monitoring & Recovery · Reasoning & Planning
  给冻结的 GPT-6 Astra 配上示范和交互记忆，用上下文适应机器人任务，无需训练机器人专用策略。
  

• RoboRSI: Stable, efficient, and reusable robot self-evolution in complex real-world environments
  2026-10-08 · Skills & Code · Self-improvement · Monitoring & Recovery · Reasoning & Planning
  把机器人自我改进组织成分层技能工程：定位负责的分支、局部修复、验证后再复用。
  

• Embodied Turing Machines: Stateful Code for Robot Recursive Self-Improvement
  2026-10-08 · Skills & Code · Memory & Knowledge · Monitoring & Recovery · Self-improvement
  让编码智能体离线改进策略库，机器人执行时只运行固定代码，以显式状态连接感知、决策和恢复。
  

• PhysEvo: Astra Can Act, Let It
  2026-10-06 · Skills & Code · Self-improvement · Monitoring & Recovery · Spatial Understanding
  冻结 Astra，让执行 Agent 行动、元 Agent 根据轨迹修订工具和技能，连诊断工具本身也能迭代。
  

• Make Code as Policy Great Again: Frontier Agents Write, Call, and Evolve Robot Tools
  2026-09-30 · Skills & Code · Reasoning & Planning · Monitoring & Recovery · Self-improvement
  把多阶段动作封装成可执行工具，每次工具运行后再由 Agent 根据反馈决定下一步。
  

• EMHO: EMbodied Agent Harness Optimization via Experience Traces
  2026-10-06 · Self-improvement · Monitoring & Recovery · Reasoning & Planning · Skills & Code
  冻结具身模型，把执行轨迹变成改进规划、感知工具和失败恢复逻辑的依据。
  

• RobotWorld: Benchmarking Multimodal Agents for Robot Use Across Diverse Tasks and Embodiments
  2026-10-07 · Benchmarks & Evaluation · Spatial Understanding · Monitoring & Recovery
  An 84-task simulation benchmark of multimodal robot-use agents across five domains, with trajectory-level diagnosis of state tracking and recovery failures.
  

• Are Frontier VLM Agents Ready to Be Robot Generalists? An Empirical Study with the Embodied Agent Arena
  2026-10-01 · Benchmarks & Evaluation · Spatial Understanding · Reasoning & Planning
  用几何、空间关系、可供性、规划和操作五个维度，检查前沿 VLM 的局部优势能否转化为完整任务成功。
  

• LIBERO-Agent: Evaluating General-Purpose Agents for Direct Embodied Manipulation
  2026-09-30 · Benchmarks & Evaluation · Reasoning & Planning · Monitoring & Recovery
  让通用 Agent 自己选择观测、处理图像和发出原生机器人命令，拆开评估感知、短程动作与长程组合。
  

• An Unexpected Robot Policy: Early Evaluations of GPT-6 Astra on RoboDojo and Beyond
  2026-09-21 · Benchmarks & Evaluation · Reasoning & Planning · VLA & World Models
  系统测试 GPT-6 Astra 直接充当机器人策略的能力，揭示语义泛化与精细动作之间的明显差距。
  

Project previews; collection dates are labeled when publication dates are unverified.

• Physical RSI 1.0: Recursive Self-Harness for Scaling Embodied Skills
  added 2026-10-12 · Project preview
  A preview system for embodied self-harness: propose changes to skills, memory and tools, evaluate frozen candidates on fresh paired cases, select a survivor and preserve evidence and lineage.
  

• LoopROS: A Recursively Self-Evolving Robot Operating System for Agentic Manipulation
  added 2026-10-12 · Project preview
  An agentic robot runtime combining within-episode observation and action feedback with background consolidation of reusable cross-episode Skills. The official project reports 82/100 successes on a selected ten-task simulation subset; component ablations are ongoing.
  

Author perspectives and essays, listed separately from experimental papers.

• Robot-Use Agents
  2026-09-07 · Phillip Isola · Perspective / essay
  Phillip Isola considers robot use as a tool capability of general-purpose agents and its implications for how capabilities spread across robots.
  

Suggestions and corrections are welcome via Issues or pull requests. Please include the paper link, first publication date, author project/code links when available, and a short explanation of its relevance to embodied agents. See contribution guidance.

Paper figures and paper content belong to their respective authors. Please cite the original work. BibTeX exports are reading aids; verify complete author names and publication metadata before submission.