{"paper":{"title":"An Embodied Generalist Agent in 3D World","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"LEO trains as a 3D embodied generalist agent through two-stage alignment on large vision-language and vision-language-action datasets.","cross_cats":["cs.AI","cs.CL","cs.LG"],"primary_cat":"cs.CV","authors_text":"Baoxiong Jia, Jiangyong Huang, Puhao Li, Qing Li, Silong Yong, Siyuan Huang, Song-Chun Zhu, Xiaojian Ma, Xiongkun Linghu, Yan Wang","submitted_at":"2023-11-18T01:21:38Z","abstract_excerpt":"Leveraging massive knowledge from large language models (LLMs), recent machine learning models show notable successes in general-purpose task solving in diverse domains such as computer vision and robotics. However, several significant challenges remain: (i) most of these models rely on 2D images yet exhibit a limited capacity for 3D input; (ii) these models rarely explore the tasks inherently defined in 3D world, e.g., 3D grounding, embodied reasoning and acting. We argue these limitations significantly hinder current models from performing real-world tasks and approaching general intelligenc"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Through extensive experiments, we demonstrate LEO's remarkable proficiency across a wide spectrum of tasks, including 3D captioning, question answering, embodied reasoning, navigation and manipulation.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The central claim assumes that the collected large-scale 3D VL and VLA datasets plus the two-stage training procedure are sufficient to produce generalist performance that transfers beyond the specific benchmarks shown.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"LEO is an embodied generalist agent that performs 3D captioning, question answering, reasoning, navigation, and manipulation after 3D vision-language alignment followed by vision-language-action instruction tuning on large-scale object- and scene-level datasets.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"LEO trains as a 3D embodied generalist agent through two-stage alignment on large vision-language and vision-language-action datasets.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"1ba28f9ab79b4d87813e61ea5758858db44d6e974a02ee779f4e777be0d75666"},"source":{"id":"2311.12871","kind":"arxiv","version":3},"verdict":{"id":"d9ad041c-600a-47ef-85dc-ec0ba0765a0a","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T14:15:53.121198Z","strongest_claim":"Through extensive experiments, we demonstrate LEO's remarkable proficiency across a wide spectrum of tasks, including 3D captioning, question answering, embodied reasoning, navigation and manipulation.","one_line_summary":"LEO is an embodied generalist agent that performs 3D captioning, question answering, reasoning, navigation, and manipulation after 3D vision-language alignment followed by vision-language-action instruction tuning on large-scale object- and scene-level datasets.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The central claim assumes that the collected large-scale 3D VL and VLA datasets plus the two-stage training procedure are sufficient to produce generalist performance that transfers beyond the specific benchmarks shown.","pith_extraction_headline":"LEO trains as a 3D embodied generalist agent through two-stage alignment on large vision-language and vision-language-action datasets."},"references":{"count":27,"sample":[{"doi":"","year":2022,"title":"A Multitask, Multilingual, Multimodal Evaluation of ChatGPT on Reasoning, Hallucination, and Interactivity","work_id":"41dff74c-00b2-4c77-a674-9f86030c06c8","ref_index":1,"cited_arxiv_id":"2302.04023","is_internal_anchor":true},{"doi":"","year":2022,"title":"RT-1: Robotics Transformer for Real-World Control at Scale","work_id":"e11bda85-8531-46bc-a07f-d0ade3643ab1","ref_index":2,"cited_arxiv_id":"2212.06817","is_internal_anchor":true},{"doi":"","year":2022,"title":"Scaling Instruction-Finetuned Language Models","work_id":"8405abb1-7558-4fdf-af24-f4c52fa77a06","ref_index":3,"cited_arxiv_id":"2210.11416","is_internal_anchor":true},{"doi":"","year":2022,"title":"LLaMA-Adapter V2: Parameter-Efficient Visual Instruction Model","work_id":"0fe2cfd8-d442-4ceb-b1a9-a465704f39b2","ref_index":4,"cited_arxiv_id":"2304.15010","is_internal_anchor":true},{"doi":"","year":2001,"title":"Scaling Laws for Neural Language Models","work_id":"b7dd8749-9c45-4977-ab9b-64478dce1ae8","ref_index":5,"cited_arxiv_id":"2001.08361","is_internal_anchor":true}],"resolved_work":27,"snapshot_sha256":"637ea65950bdb3b83fba8253b8026046cca6267d4912d88a95831e9a9b795c1e","internal_anchors":10},"formal_canon":{"evidence_count":2,"snapshot_sha256":"4a0c00bfe591c510c5d6f0e5e9394e6cb682f8a0b168a0fdf54511e583e49651"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}