<a href=\"https://cdn-uploads.huggingface.co/production/uploads/64c9bced3a3f428da6902741/k3UHoPWpeGUH4Te3Z4v7R.png\" rel=\"nofollow\"><img src=\"https://cdn-uploads.huggingface.co/production/uploads/64c9bced3a3f428da6902741/k3UHoPWpeGUH4Te3Z4v7R.png\" alt=\"embodied-cpp-icon\"></a><br>We introduce <strong>Embodied.cpp</strong>, a portable <strong>C++ inference runtime for embodied AI models</strong> on heterogeneous robots.</p>\n<p>A key motivation behind this work is that the embodied AI community has focused much more on <strong>model training</strong> than on <strong>practical inference and deployment</strong>.</p>\n<p>Today, we already see many exciting embodied models, but there is still <strong>no general inference framework</strong> designed for:</p>\n<ul>\n<li><strong>closed-loop embodied inference</strong></li>\n<li><strong>edge-side / on-robot deployment</strong></li>\n<li><strong>heterogeneous devices and hardware backends</strong></li>\n<li><strong>unified support across different embodied model families</strong></li>\n</ul>\n<p>This is exactly the gap we target with <strong>Embodied.cpp</strong>: a portable <strong>C++ inference runtime</strong> for embodied AI models on heterogeneous robots.</p>\n<p>In the paper, we focus on the runtime side of embodied intelligence:</p>\n<ul>\n<li>how to support <strong>latency-first batch-1 inference</strong></li>\n<li>how to handle <strong>multi-rate closed-loop execution</strong></li>\n<li>how to bridge diverse embodied models through a unified runtime abstraction</li>\n<li>how to make deployment more practical across simulators, robots, and hardware</li>\n</ul>\n<p>We evaluate the system on representative models including <strong>HY-VLA</strong>, <strong>pi0.5</strong>, and <strong>LingBot-VA Transformer block</strong> benchmark.</p>\n<p>This is only the beginning. We plan to continue updating:</p>\n<ul>\n<li><strong>more embodied models</strong></li>\n<li><strong>more complete runtime and system support</strong></li>\n<li><strong>more edge-side and heterogeneous-device optimization methods</strong></li>\n</ul>\n<p>Code and model artifacts:</p>\n<ul>\n<li>GitHub: <a href=\"https://github.com/SEU-PAISys/Embodied.cpp\" rel=\"nofollow\">https://github.com/SEU-PAISys/Embodied.cpp</a></li>\n<li>Hugging Face: <a href=\"https://huggingface.co/SEU-PAISys/Embodied.cpp\">https://huggingface.co/SEU-PAISys/Embodied.cpp</a></li>\n</ul>\n<p>Paper: <a href=\"https://arxiv.org/abs/2607.02501\" rel=\"nofollow\">https://arxiv.org/abs/2607.02501</a></p>\n","updatedAt":"2026-07-06T03:01:02.623Z","author":{"_id":"64c9bced3a3f428da6902741","avatarUrl":"/avatars/897e149b0d661944e2d75fa708ae4b92.svg","fullname":"Borui Li","name":"libr6","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8964861631393433},"editors":["libr6"],"editorAvatarUrls":["/avatars/897e149b0d661944e2d75fa708ae4b92.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2607.02501","authors":[{"_id":"6a4a68cfd66e8ea154bc8079","name":"Ling Xu","hidden":false},{"_id":"6a4a68cfd66e8ea154bc807a","name":"Chuyu Han","hidden":false},{"_id":"6a4a68cfd66e8ea154bc807b","user":{"_id":"64c9bced3a3f428da6902741","avatarUrl":"/avatars/897e149b0d661944e2d75fa708ae4b92.svg","isPro":false,"fullname":"Borui Li","user":"libr6","type":"user","name":"libr6"},"name":"Borui Li","status":"claimed_verified","statusLastChangedAt":"2026-07-05T21:06:24.932Z","hidden":false},{"_id":"6a4a68cfd66e8ea154bc807c","name":"Hao Wu","hidden":false},{"_id":"6a4a68cfd66e8ea154bc807d","name":"Shiqi Jiang","hidden":false},{"_id":"6a4a68cfd66e8ea154bc807e","name":"Ting Cao","hidden":false},{"_id":"6a4a68cfd66e8ea154bc807f","name":"Chuanyou Li","hidden":false},{"_id":"6a4a68cfd66e8ea154bc8080","name":"Sheng Zhong","hidden":false},{"_id":"6a4a68cfd66e8ea154bc8081","name":"Shuai Wang","hidden":false}],"publishedAt":"2026-07-02T00:00:00.000Z","submittedOnDailyAt":"2026-07-06T00:00:00.000Z","title":"Embodied.cpp: A Portable Inference Runtime of Embodied AI Models on Heterogeneous Robots","submittedOnDailyBy":{"_id":"64c9bced3a3f428da6902741","avatarUrl":"/avatars/897e149b0d661944e2d75fa708ae4b92.svg","isPro":false,"fullname":"Borui Li","user":"libr6","type":"user","name":"libr6"},"summary":"Embodied AI models now span vision-language-action (VLA) models and world-action models (WAMs), but practical deployment remains fragmented across model-specific Python stacks, backend assumptions, and robot-side glue code, especially on heterogeneous edge devices. Existing inference runtimes are designed mainly for request-response serving and therefore do not satisfy the runtime contract of embodied deployment: multi-rate execution inside closed-loop control, latency-first batch-1 inference on heterogeneous hardware, and extensible embodied interfaces beyond fixed token I/O. We present Embodied.cpp, a portable C++ inference runtime for embodied models. Based on an architectural analysis of representative VLA models and WAMs, Embodied.cpp captures a shared execution path and organizes it into five layers: input adapters, sequence builders, backbone execution, head plugins, and deployment adapters. The runtime provides modular multi-rate execution, latency-first fused inference, and extensible operator and I/O support, enabling deployment across heterogeneous devices, robots, and simulators through one backend abstraction. We evaluate Embodied.cpp on two VLA models, HY-VLA and pi0.5, and on a preliminary WAM benchmark using a LingBot-VA Transformer block. The VLA deployments achieve successful closed-loop execution with 100.0% and 91.0% task success rates, respectively. The WAM benchmark reduces block memory from 312.2 MiB to 88.1 MiB. These results show that Embodied.cpp improves deployment efficiency while preserving high accuracy across diverse embodied model architectures.","upvotes":37,"discussionId":"6a4a68cfd66e8ea154bc8082","githubRepo":"https://github.com/SEU-PAISys/Embodied.cpp","githubRepoAddedBy":"user","ai_summary":"Embodied.cpp is a portable C++ runtime that enables efficient deployment of vision-language-action and world-action models across heterogeneous edge devices through modular execution layers and optimized inference.","ai_keywords":["Vision-language-action models","world-action models","inference runtime","closed-loop control","multi-rate execution","latency-first inference","fused inference","embodied interfaces","hardware heterogeneity","model deployment"],"ai_summary_model":"Qwen/Qwen2.5-Coder-32B-Instruct","githubStars":61,"organization":{"_id":"6a4a1ed8d0349cf4af0141f0","name":"SEU-PAISys","fullname":"PhysicalAI System Group @ SAIL Lab, Southeast University","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/64c9bced3a3f428da6902741/rVDcprqI4aVzfGcujJfn-.jpeg"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"64c9bced3a3f428da6902741","avatarUrl":"/avatars/897e149b0d661944e2d75fa708ae4b92.svg","isPro":false,"fullname":"Borui Li","user":"libr6","type":"user"},{"_id":"6407e5294edf9f5c4fd32228","avatarUrl":"/avatars/8e2d55460e9fe9c426eb552baf4b2cb0.svg","isPro":false,"fullname":"Stoney Kang","user":"sikang99","type":"user"},{"_id":"620783f24e28382272337ba4","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/620783f24e28382272337ba4/zkUveQPNiDfYjgGhuFErj.jpeg","isPro":false,"fullname":"GuoLiangTang","user":"Tommy930","type":"user"},{"_id":"642653df22bddcea3d284b4d","avatarUrl":"/avatars/2a328c8f1fe5dc30bef59b6bd6dc4a79.svg","isPro":false,"fullname":"Hao Wu","user":"HakoWu","type":"user"},{"_id":"662852bf8bfc90408a97acdf","avatarUrl":"/avatars/008e6faf51e1be4f0c0d17008dfe2f49.svg","isPro":false,"fullname":"cao","user":"tingcao","type":"user"},{"_id":"65f29e6e2b802e87131dc51e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/afEAwAe5qdCj8rQMo5-J6.jpeg","isPro":false,"fullname":"NIME","user":"Nime2568","type":"user"},{"_id":"6a4b60f55a0a9ecbdb0d4f1e","avatarUrl":"/avatars/19e2cd8795dca529836335bf55621d83.svg","isPro":false,"fullname":"jiameilv","user":"jiameilv","type":"user"},{"_id":"66b3acba55607725488b3fb3","avatarUrl":"/avatars/5e223272fc59c21834760863aefc4686.svg","isPro":false,"fullname":"xiao.li","user":"m1rr0r1999","type":"user"},{"_id":"67c8806d4c6d0ed05625168f","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/BByAH406ylob6FZIP2WU0.png","isPro":false,"fullname":"Shimao Xu","user":"xuwto","type":"user"},{"_id":"6905c9b238993e5e3f8ae505","avatarUrl":"/avatars/cbbefc40832ae1f4d31ed6a33b9eeded.svg","isPro":false,"fullname":"郭耀宇","user":"gyybalding","type":"user"},{"_id":"699e9c5b9b93bc6afbfabbf7","avatarUrl":"/avatars/aca99ea58c13c506000c67953ebfb8a0.svg","isPro":false,"fullname":"Taylor Levi","user":"donghao56","type":"user"},{"_id":"63ac5701c21e60a3e9b58aa7","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63ac5701c21e60a3e9b58aa7/g6EX7diOpuA94R2ab-rZC.png","isPro":true,"fullname":"Dipankar Sarkar","user":"dipankarsarkar","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":2,"organization":{"_id":"6a4a1ed8d0349cf4af0141f0","name":"SEU-PAISys","fullname":"PhysicalAI System Group @ SAIL Lab, Southeast University","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/64c9bced3a3f428da6902741/rVDcprqI4aVzfGcujJfn-.jpeg"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2607/2607.02501.md","query":{}}">

Get this paper in your agent:

Don't have the latest CLI?

Cite arxiv.org/abs/2607.02501 in a dataset README.md to link it from this page.

Cite arxiv.org/abs/2607.02501 in a Space README.md to link it from this page.

Add this paper to a collection to link it from this page.