Embodied AI Radar · 具身智能雷达

A bilingual (English / 中文), always-fresh map of embodied AI and robot learning. Hand-verified papers, code, datasets, simulators and hardware with one-line takeaways — plus every new robotics paper on arXiv, fetched and tagged by topic every day.

🌐 Open the website → hu-zlatan.github.io/embodied-ai-radar

CoRL 2026 (Nov 9–12, Austin): the program guide lists all 687 accepted papers by session and poster number, with filters and arXiv/code links, plus the 32 workshops.

New to the field? Follow the reading path: 10 steps from RL and imitation basics through visuomotor policies, VLA models, simulation, locomotion and world models to hardware you can build — 41 hand-picked entries, each with a bilingual takeaway.

Why this one

• Verified. Every curated arXiv ID is checked against arXiv and every link is checked by CI, weekly.
• Fresh. A GitHub Action pulls new cs.RO papers (plus robotics papers from cs.CV / cs.LG / cs.AI / cs.CL) every day and tags them by topic — they appear on the site without anyone touching the repo.
• Searchable. Instant search and filters (topic · robot type: arm, hand, humanoid, legged, drone… · accepted venue such as CoRL, NeurIPS, IROS · curated vs. daily · has code) on a fast static site that works on phones. Papers accepted at a venue also get a page per conference, e.g. CoRL 2026 (all venues), and every robot type has its own page and Atom feed, e.g. humanoid · dexterous hand · quadruped. You can also browse by lab or company, e.g. NVIDIA · Stanford · Shanghai AI Lab (all labs).
• Signal, not just volume. Each paper shows the GitHub stars of its code and, when it made Hugging Face Daily Papers, its community upvotes (▲) — sort search by either, or see the most-upvoted robotics papers of the last two weeks on the home page.
• Code tracker. Every curated repo on one page: GitHub stars, stars gained this week, last update and detected license, with filters by topic and license.

What	How often
Daily arXiv feed (site · Atom; one Atom feed per topic, e.g. VLA)	every day, automatically
Curated index (below and on the site) · agent picks from the daily feed (Atom, a few a day, bilingual)	several times a week
Weekly digest	weekly

Topics

VLA & robot foundation models · manipulation & dexterous hands · tactile sensing · locomotion & humanoids · world models · navigation & embodied agents · perception, SLAM & 3D mapping · planning, control & safety · sim-to-real & simulators · datasets & benchmarks · hardware & open-source robots.

Curated index

Editor's rating: ⭐⭐⭐ must-read (at most one per topic) · ⭐⭐ highly recommended · ⭐ recommended · no star = worth knowing. Each topic lists the rated entries first, then the newest picks; All N entries opens the complete list. Generated from data/curated/ — edit the YAML, not these lists.

VLA & Robot Foundation Models · Manipulation & Dexterous Hands · Tactile & Visuotactile Sensing · Locomotion & Humanoids · World Models · Navigation & Embodied Agents · Perception, SLAM & 3D Mapping · Planning, Control & Safety · Robot Learning: RL & Imitation · Sim-to-Real & Simulators · Datasets & Benchmarks · Hardware & Open-Source Robots

VLA & Robot Foundation Models

Vision-language-action models and generalist robot policies. Browse with the latest papers →

• ⭐⭐⭐ π0: A Vision-Language-Action Flow Model for General Robot Control (2024) — A VLM backbone with a flow-matching action expert, trained across many robots, doing long dexterous tasks like laundry folding. · code
• ⭐⭐ A Pragmatic VLA Foundation Model (2026) — LingBot-VLA: an Apache-2.0 foundation model trained on about 20,000 hours of real data from nine dual-arm setups, tested on four platforms with 100 tasks each; code, base model and benchmark data are released. · code · project
• ⭐⭐ From Pixels to Actions - Towards Native Vision-Language-Action Policies (2026) — NEO-VLA drops the visual encoder: images and language evolve in one backbone and a flow-matching action expert reads several depths. Layer probing shows action signals often peak before the final visual layer. · code · project
• ⭐⭐ Open-HumanWAM: An Open Recipe for Scaling Human and Robot Experience for Embodied Foundation Models (2026) — Opens a 50,000-hour annotated egocentric video corpus with pipelines and checkpoints. Reports a log-linear law between pretraining hours and robot success, and 84.6% on RoboTwin 2.0 Clean2Random. · code · project
• ⭐⭐ RoboColosseum: Evaluating Physical Intelligence at Scale In-The-Wild (2026) — Blinded pairwise trials at 124 sites compare VLAs and frontier agents. Rankings barely agree across robots (W = 0.24), and open rankings poorly predict results after fine-tuning; no policy finishes a G1 task. ·

All 258 entries →

Manipulation & Dexterous Hands

Imitation learning, diffusion policies, grasping, bimanual and dexterous manipulation. Browse with the latest papers →

• ⭐⭐⭐ Diffusion Policy: Visuomotor Policy Learning via Action Diffusion (2023) — Generating action sequences by denoising handles multimodal demos far better than regression; now a standard IL baseline. · code
• ⭐⭐ AIRoA-0: Long-Horizon, Precise, Cross-Embodiment Manipulation from Handheld Data (2026) — Learned from 16,151 hours of handheld gripper demos with no robot data; reports 78% average success over ten tasks versus 27% for the best other VLA, and releases the dataset, weights and post-training recipe. · code · project
• ⭐⭐ Generalizable Humanoid Manipulation with 3D Diffusion Policies (2024) — iDP3: whole-upper-body teleop, a 25-DoF humanoid with 3D LiDAR, and an improved 3D Diffusion Policy for noisy human data, aiming at skills beyond a single scene. · code · project
• ⭐⭐ Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots (2024) — A handheld gripper with a GoPro collects demos anywhere; policies transfer directly to robot arms. · code
• ⭐⭐ Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (2023) — ALOHA + ACT — a ~$20k bimanual teleop rig and an action-chunking Transformer that learns precise tasks from ~50 demos. · code

All 170 entries →

Tactile & Visuotactile Sensing

Tactile sensors, touch representations and contact-rich control. Browse with the latest papers →

• ⭐⭐⭐ DIGIT: A Novel Design for a Low-Cost Compact High-Resolution Tactile Sensor with Application to In-Hand Manipulation (2020) — An inexpensive, compact GelSight-style fingertip sensor that made vision-based touch widely accessible. · code
• ⭐ Visible Touch: Rendering Contact for Visuomotor Policies (2026) — CoRL 2026 Spotlight. Brings contact into image-conditioned policies by rendering it, targeting the gap left by approaches that need tactile hardware, separate tactile encoders or non-image backbones. · code · project
• Blind Dexterous Grasping via Real2Sim2Real Tactile Policy Learning (2026) — Tactile-only blind grasping on a LEAP Hand: calibrate a digital twin to reproduce real touch signals, pretrain a layout-aware tactile encoder, and distill object-specific RL experts into a diffusion policy.
• Equivariant Visual-Tactile Diffusion Policy for Contact-Rich Manipulation (2026) — VISTA: equivariant visuotactile diffusion policy that fuses touch into spherical visual tokens and rotates them with the gripper pose, for more data-efficient contact-rich imitation. · code · project
• FM-VLA: Force-based Memory for Vision-Language-Action Models in Contact-Rich Manipulation (2026) — Gives a VLA memory through compact force-history tokens from a pretrained VAE, for tasks like pressing a button a set number of times. Reports over 80% success on three memory-dependent tasks. · project

All 76 entries →

Locomotion & Humanoids

Legged locomotion, humanoid whole-body control and parkour. Browse with the latest papers →

• ⭐⭐⭐ Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning (2021) — Thousands of parallel GPU environments train a quadruped walking policy in minutes; legged_gym became a standard starter. · code
• ⭐⭐ Perceptive Humanoid Motion from Simulation (2026) — HALO trains humanoid skills on vision-language-motion data synthesised in a thousand procedural homes, so data scales with compute. Reports zero-shot transfer: 91% over 360 trials in six real environments. · code · project
• ⭐⭐ BeyondMimic: From Motion Tracking to Versatile Humanoid Control via Guided Diffusion (2025) — One motion-tracking recipe with shared hyperparameters learns agile skills (cartwheels, flip-kicks); a guided latent diffusion model composes them for new goals like joystick control, zero-shot on hardware. · code
• ⭐⭐ SONIC: Supersizing Motion Tracking for Natural Humanoid Whole-Body Control (2025) — Tests scaling on humanoid control: more network, data and compute yield a generalist motion-tracking foundation model, with no manual reward engineering.
• ⭐⭐ Real-World Humanoid Locomotion with Reinforcement Learning (2023) — A causal Transformer policy trained in simulation walks a full-size humanoid outdoors, zero-shot.
• ⭐⭐ DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills (2018) — Showed RL can imitate a wide range of motion clips, flips and spins included, while staying robust. The motion-imitation template much humanoid work builds on. ·

All 170 entries →

World Models

Learned dynamics, video world models, world action models (WAM) and model-based RL for embodied agents. Browse with the latest papers →

• ⭐⭐⭐ Mastering Diverse Domains through World Models (2023) — DreamerV3 — one set of hyperparameters works across 150+ tasks, including collecting diamonds in Minecraft from scratch. · code
• ⭐⭐ Cosmos 3: Omnimodal World Models for Physical AI (2026) — A family of omnimodal models that handles language, image, video, audio and action in one mixture-of-transformers, folding VLMs, video generators, world simulators and world-action models into a single framework. · code · project
• ⭐⭐ DreamDojo: A Generalist Robot World Model from Large-Scale Human Videos (2026) — A foundation world model pretrained on 44k hours of egocentric human video, using continuous latent actions as proxies for missing labels, then post-trained on small robot datasets. · code · project
• ⭐⭐ OpenWAM: An Open, Modular Exploration Towards Systematic World-Action Model Pretraining (2026) — An open research stack that splits world-action models into composable modules, then runs controlled studies on what to inherit from video priors and how world and action learning interact. · code · project
• ⭐⭐ RoboJEPA: Scaling Robotic Latent World Models (2026) — A JEPA world model trained across 12 robot embodiments; its imagination error follows a power law in compute, so model quality can be predicted beyond the fitted scale. · ·

All 128 entries →

Navigation & Embodied Agents

Vision-language navigation, embodied question answering, mobile manipulation and LLM planners. Browse with the latest papers →

• ⭐⭐⭐ Do As I Can, Not As I Say: Grounding Language in Robotic Affordances (2022) — SayCan — an LLM proposes steps and learned value functions score what the robot can actually do right now.
• ⭐ Driving, Fast or Slow? Neuro-Symbolic Guidance for Motion Prediction in Multi-Modal Ground Mobility (2026) — TraCS adds interpretable, probabilistic traffic-rule logic to black-box motion predictors, generated from natural-language regulations; reports consistent gains on Argoverse 2. Oral.
• ⭐ ViNT: A Foundation Model for Visual Navigation (2023) — A Transformer navigation policy trained on many robots' driving data that adapts to new goal types and robots with light fine-tuning. · code
• ⭐ GNM: A General Navigation Model to Drive Any Robot (2022) — The predecessor of ViNT — one goal-conditioned navigation policy trained on pooled data from several different robots, which transfers to new robots including ones not seen in training. · code
• ⭐ The Marathon 2: A Navigation System (2020) — The paper behind Nav2, the standard ROS 2 navigation stack - behavior-tree orchestration of planners, controllers and recovery behaviors, validated by long marathon runs on real robots. · code
• AcroRL: Learning Aggressive Quadrotor Inversion using Bidirectional Thrust (2026) — RL policies modulate a reference trajectory for compact quadrotor flips with reversible thrust, one policy per direction. In JAX simulation: 32% lower position RMSE and 57% shorter settling than baselines.
• Adaptive Risk-Certified Event-Triggered Replanning for Dynamic Navigation (2026) — Uses calibrated barrier risk (CVaR along the MPC rollout) as an early warning to switch to a safer planning mode. Reports 83.3% fewer collisions than a safety-filter-only baseline. ·

All 83 entries →

Perception, SLAM & 3D Mapping

Visual / LiDAR SLAM and odometry, 3D reconstruction, object pose, depth and open-vocabulary scene maps for robots. Browse with the latest papers →

• ⭐⭐⭐ ORB-SLAM3: An Accurate Open-Source Library for Visual, Visual-Inertial and Multi-Map SLAM (2020) — The go-to feature-based SLAM library — monocular, stereo and RGB-D, with IMU fusion and multi-map relocalization. · code
• ⭐⭐ Segment Anything (2023) — SAM — a promptable segmentation model (points, boxes, masks) trained on 1B+ masks; a standard building block for object masks in robot perception pipelines. · code
• ⭐ FoundationStereo: Zero-Shot Stereo Matching (2025) — Stereo depth model trained on large synthetic data that reports strong zero-shot accuracy on real scenes — a drop-in depth source for stereo-camera robots. · code
• ⭐ SAM 3: Segment Anything with Concepts (2025) — Segments and tracks every instance of a concept given a short noun phrase or image exemplar, in images and video; trained with a data engine covering 4M concept labels, with open weights and the SA-Co benchmark. · code
• ⭐ VGGT: Visual Geometry Grounded Transformer (2025) — One feed-forward transformer predicts cameras, depth, point maps and tracks for many views at once, in about a second. · code
• ⭐ Depth Anything V2 (2024) — Trains a large teacher only on synthetic images with precise depth, then pseudo-labels many real images to train smaller students; gives sharper, more robust monocular depth than V1. · code

All 53 entries →

Planning, Control & Safety

Motion planning, trajectory optimization, MPC, safety filters and safe learning-based control. Browse with the latest papers →

• ⭐⭐⭐ cuRobo: Parallelized Collision-Free Minimum-Jerk Robot Motion Generation (2023) — GPU-parallel motion generation for arms — many trajectory optimizations run at once, giving collision-free, smooth plans in milliseconds. · code
• ⭐ cuRoboV2: Dynamics-Aware Motion Generation with Depth-Fused Distance Fields for High-DoF Robots (2026) — cuRoboV2 adds torque-aware B-spline trajectory optimization and GPU-native TSDF/ESDF perception for high-DoF robots. · code
• ⭐ PyRoki: A Modular Toolkit for Robot Kinematic Optimization (2025) — PyRoki: a modular, cross-platform kinematic optimization toolkit that runs natively on CPU, GPU and TPU. · code · project
• ⭐ Planning with Diffusion for Flexible Behavior Synthesis (2022) — Diffuser — treats planning as denoising whole trajectories, so goals and rewards can steer sampling; an early link between diffusion and control. · code
• ⭐ Predictive Sampling: Real-time Behaviour Synthesis with MuJoCo (2022) — Introduces MuJoCo MPC (MJPC), an interactive tool for real-time predictive control in MuJoCo, and shows that a simple sampling-based planner is a strong baseline for contact-rich behaviours. · code
• ⭐ legged_control (2022) — Open-source NMPC + whole-body control + state-estimation stack for quadrupeds built on OCS2 and ros-control, with a sim-to-real path; a popular starting point for model-based legged control.

All 50 entries →

Robot Learning: RL & Imitation

Reinforcement learning, imitation learning and policy post-training methods for robots, on any embodiment. Browse with the latest papers →

• ⭐⭐⭐ Proximal Policy Optimization Algorithms (2017) — The clipped policy-gradient method behind most GPU-parallel sim-to-real locomotion and dexterity training.
• ⭐⭐ Precise and Dexterous Robotic Manipulation via Human-in-the-Loop Reinforcement Learning (2024) — HIL-SERL — real-world RL with human corrections, reported to reach near-perfect success on precise assembly and dynamic tasks after a few hours of training. · code
• ⭐ Control-Ready Uncertainty for Trajectory Diffusion (2026) — SCOPE gives diffusion trajectory models calibrated uncertainty tubes without Monte Carlo sampling, by distilling score curvature; tested in crowd navigation and Franka manipulation. · project
• ⭐ Diffusion ReRoll: Revisable Denoising for Robotic Sequential Prediction (2026) — CoRL 2026 Spotlight. Diffusion sequence prediction with selective re-noising, so segments of the horizon can be refined again using context from the rest. · project
• ⭐ FedGuide: Diffusion Prior Alignment and Value Baseline Guidance for Heterogeneous Federated Reinforcement Learning (2026) — Federated RL for heterogeneous clients that aggregates diffusion behavior priors with optimal transport rather than averaging policies, plus a low-variance DICE value baseline.
• ⭐ Learning from Fleet Experience: Collective Experience Learning Across 96 Robots (2026) — Deployed robots pool their runs; failures retrieve matching successful segments from the fleet buffer to fine-tune on, with no new demos. Reports 93–99% on five real tasks; 96 robots lift phone packing 79% → 98%. · code ·

All 132 entries →

Sim-to-Real & Simulators

Physics simulators, GPU-parallel training and transfer to real robots. Browse with the latest papers →

• ⭐⭐⭐ MuJoCo (2012) — Fast, accurate contact physics; the default simulator for robot learning research, open-sourced by DeepMind. · code
• ⭐⭐ When the Simulator Fails, Not the Policy: Auditing Contact Artifacts in Simulated Robot Benchmarks (2026) — Many benchmark 'policy failures' are physics-engine ejections. SimuGuard traced RoboTwin artifacts to one collision constant; changing it cut artifact episodes from 12/100 to none and lifted five tasks 89% → 98%. · code
• ⭐ Genie Sim 3.0 : A High-Fidelity Comprehensive Simulation Platform for Humanoid Robot (2026) — Genie Sim 3.0: a manipulation simulation platform whose LLM tool builds scenes from natural language, with an LLM- and VLM-based automated benchmark. · code
• ⭐ Isaac Lab: A GPU-Accelerated Simulation Framework for Multi-Modal Robot Learning (2025) — The reference paper for Isaac Lab, Orbit's successor — GPU-parallel physics, rendering and sensor simulation with ready-made robots and tasks for RL and imitation. · code
• ⭐ MuJoCo Playground (2025) — GPU-accelerated MJX environments for locomotion and manipulation that train in minutes and transfer to real robots. · code
• ⭐ Newton (2025) — Open, GPU-accelerated physics engine built on NVIDIA Warp and OpenUSD with multiple solvers for contact-rich robot learning; co-developed by Disney Research, Google DeepMind and NVIDIA, hosted by the Linux Foundation.

All 63 entries →

Datasets & Benchmarks

Large robot datasets, egocentric data and evaluation suites. Browse with the latest papers →

• ⭐⭐⭐ Open X-Embodiment: Robotic Learning Datasets and RT-X Models (2023) — 1M+ trajectories from 22 robot types pooled by 30+ labs; RT-X models trained on it beat single-robot baselines. · code
• ⭐⭐ LeRobot (2024) — PyTorch library with datasets, pretrained policies (ACT, Diffusion, SmolVLA) and drivers for low-cost arms.
• ⭐⭐ LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning (2023) — 130 language-conditioned tasks in four suites; now the most common simulation benchmark for VLA papers. · code
• ⭐ BEAKER: An Expert-Curated Benchmark for Embodied Brains in Self-Driving Chemical Laboratories (2026) — 1,000 expert-curated questions over 11 image and video task types test multimodal LLMs as the brain of a self-driving chemistry lab: affordances, trajectories, apparatus setup, safety and step reasoning.
• ⭐ WireCraft: A Simulation Benchmark for Industrial DLO Manipulation (2026) — CoRL 2026 Spotlight. Simulation benchmark for wires and cables in industrial assembly, aimed at the gap left by benchmarks tied to specific hardware or lacking modular, customizable task assets.
• ⭐ RoboVerse: Towards a Unified Platform, Dataset and Benchmark for Scalable and Generalizable Robot Learning (2025) — MetaSim puts many simulators (Isaac Sim/Gym, MuJoCo, Genesis, SAPIEN…) behind one interface; adds a large synthetic dataset migrated from public benchmarks and unified IL/RL benchmarks. · code · project
• ⭐ (2024) — SimplerEnv — real-to-sim environments whose scores correlate with real-robot results for RT-1/Octo-style policies. ·

All 97 entries →

Hardware & Open-Source Robots

Low-cost arms, hands, grippers and data-collection rigs you can build. Browse with the latest papers →

• ⭐⭐⭐ LEAP Hand: Low-Cost, Efficient, and Anthropomorphic Hand for Robot Learning (2023) — A 16-DoF dexterous hand buildable for about $2,000 from off-the-shelf parts. · code · project
• ⭐⭐ Pixel: A Co-Designed Full-Stack Robot System for Learning General-Purpose Mobile Manipulation (2026) — Hardware, data and learning designed together: a low-cost wheeled humanoid with compliant grippers, teleoperation and robot-free human data collection, one whole-body controller and a shared policy pipeline. · project
• ⭐ Demonstrating Berkeley Humanoid Lite: An Open-source, Accessible, and Customizable 3D-printed Humanoid Robot (2025) — Open-source humanoid with 3D-printed modular gearboxes, buildable for under US$5,000; hardware designs, embedded code and training code are released. · code · project
• ⭐ OpenArm (2025) — Fully open-source 7-DoF humanoid arm for contact-rich work — CAD, electronics, firmware, ROS 2 and MuJoCo/Genesis models, plus leader–follower force-feedback teleop and gravity compensation. · project
• ⭐ ALOHA 2: An Enhanced Low-Cost Hardware for Bimanual Teleoperation (2024) — Upgraded ALOHA bimanual rig (stronger grippers, gravity compensation, sturdier frame) with open hardware designs and a MuJoCo model for simulation. · code · project

All 51 entries →

Related lists & learning resources

Other good maps of the field, plus courses and docs we point newcomers to. This radar complements them with search, a daily feed and bilingual takeaways.

Related lists

• Embodied-AI-Guide 中文 — A large Chinese-language beginner's guide to embodied AI, from fundamentals to research directions.
• Embodied_AI_Paper_List (HCPLab-SYSU) — Paper list accompanying a survey of embodied AI, organized by perception, interaction, agents and sim-to-real.
• awesome-humanoid-robot-learning — A focused list of humanoid robot learning papers.
• Awesome-Robotics-Foundation-Models — Papers on foundation models in robotics, from the survey "Foundation Models in Robotics".
• awesome-embodied-vla-va-vln — A frequently updated list of vision-language-action (VLA) and vision-language navigation (VLN) papers.

Courses & textbooks

• Robotic Manipulation (MIT 6.4210, Russ Tedrake) — Free online textbook and course on perception, planning and control for manipulation, with Drake notebooks.
• Underactuated Robotics (MIT, Russ Tedrake) — Free online textbook on nonlinear dynamics, optimal control and trajectory optimization for walking, flying and swimming robots.
• CS 285 Deep Reinforcement Learning (UC Berkeley) — Sergey Levine's deep RL course; lecture videos and slides are public.
• 蘑菇书 EasyRL (Datawhale) 中文 — A popular open Chinese reinforcement-learning tutorial (from Q-learning to PPO, DDPG and TD3) with code; readable online.
• 动手学强化学习 (Hands-on RL) 中文 — Chinese RL textbook built from runnable Jupyter notebooks; the online version is at hrl.boyuai.com.

Tool docs

• LeRobot documentation — Hands-on docs for Hugging Face LeRobot — record data, train ACT/diffusion/VLA policies and run them on low-cost arms.
• Isaac Lab documentation — Docs and tutorials for NVIDIA's GPU-parallel robot learning framework.
• MuJoCo documentation — Reference and modeling guide for the MuJoCo physics engine (MJX, MJCF, Python bindings).

Communities

• 古月居 (GuYueHome) 中文 — A long-running Chinese ROS and robotics community with tutorials and forums.

Contribute

Found a great paper, repo, dataset or robot that's missing — or a mistake? Open an issue or send a pull request that adds one entry to data/curated/<topic>.yaml. See CONTRIBUTING.md.

How this project is run

This repository is built and maintained by an AI agent (Claude, made by Anthropic) that works on a schedule, under the oversight of @hu-zlatan. The agent's brief is AUTOPILOT.md; its notes, decisions and run log are in .autopilot/. Data is never invented: entries that fail verification are fixed or removed.

License

Code: MIT. Curated text and data: CC BY 4.0. arXiv metadata: CC0, from the arXiv API — thank you to arXiv for use of its open access interoperability.