论文记录 · 证据边界公开

NaVILA: Legged Robot Vision-Language-Action Model for Navigation

Robotics: Science and Systems 2025 · 2024-12-05 · 空间感知与导航

论文元数据

作者: An-Chieh Cheng、Yandong Ji、Zhaojing Yang、Zaitian Gongye、Xueyan Zou、Jan Kautz、Erdem Bıyık、Hongxu Yin、Sifei Liu、Xiaolong Wang
机构: University of California, San Diego、NVIDIA
发布日期精度: 日
研究方向: 空间感知与导航
机器人: 未关联具体机器人
DOI: 10.15607/RSS.2025.XXI.018
arXiv: 2412.04453
arXiv 当前版本: v2
arXiv 最近修订: 2025-02-17
arXiv 主分类: cs.RO
引用元数据: 0
记录状态: included
质量检查: 无自动质量红旗

工程资源状态

代码: 已核验开放
权重: 存在链接线索，尚未核验官方性与可运行性
数据: 存在链接线索，尚未核验官方性与可运行性
许可证: 部分制品许可证缺失/待澄清

当前实践优先级

高潜待持续验证 · 81/100 · 13–24 个月

路线标签：可能成为新主流、足式VLA导航范式转折

社区证据类型：社区正向/混合验证

NaVILA把慢速语言视觉规划与高速足式运动分层，是当前开放词汇腿式导航最重要的可部署参考之一；Go2/T1实机、代码、权重、数据和独立Go2集成都存在，但许可证与独立成功率仍未闭环。

近期窗口 · 28/35
发布距评分基准日 19 个月，落入 13–24 个月窗口。
社区可复查验证 · 15/25
存在独立真实Go2系统集成和完整依赖说明，但没有与论文同口径的量化复现。
跨领域技术创新 · 20/20
VLM长时语义、离散语言动作、高速运动策略、足式导航和跨本体数据连接。
量化实机证据 · 10/10
Unitree Go2与Booster T1跨办公、家庭、户外的简单/复杂指令均有成功率。
开放与可复现 · 3/5
代码/权重/约46.16GB数据/benchmark可取，但权重和数据许可缺失。
工程实用增量 · 5/5
分层架构把慢VLM与快控制解耦，并披露FP16/W4A16时延显存。

社区可复查证据：

社区证据 1
独立BSD-2 ROS2系统固定提交接入NaVILA、Go2、8帧缓存、里程计/雷达安全闭环和Blackwell修复，但未给任务成功率分母。

证据缺口：权重与数据许可缺失；FP16 RTX4090约594.58ms/18.5GB；独立集成无成功率分母；原始媒体权利需核对；复杂开放世界安全仲裁不足

当前实践分只决定今天的阅读与最小验证优先级；近期、社区验证和跨领域创新合计 80/100，不会把前沿观察条目自动升级成候选或确认里程碑。社区维度只有显式独立证据才能得高分，只有失败记录时必须标为社区反证。

已终审工程资源

官方代码 · 许可证 Apache-2.0 · 最近核验 2026-07-20
依赖：Python 3.10、VILA training stack、FlashAttention 2、Transformers 4.37.2、VLN-CE and Habitat evaluation dependencies、WebDataset 0.1.103；硬件/传感器：One or more CUDA GPUs for the documented evaluation path、Legged robot hardware for physical navigation; exact SDK boundary is not fully packaged in the README；实机证据：The paper/project reports physical legged-robot navigation; the repository also links weights, datasets and an Isaac Sim benchmark, but this batch promotes code only.；开放范围/限制：Verified at commit 76b98f233dd0fff05dfcd69435eec6740febff9d. Model weights and datasets must be licensed and verified separately; Apache-2.0 is not automatically inherited by them.
官方代码 · 许可证 MIT for repository code at commit e9d2db12ce5788c0f987d734c0094100b6bc0d3a; Isaac Sim, scene assets and three acknowledged upstream benchmark repositories retain their own terms · 最近核验 2026-07-22
依赖：NVIDIA Isaac Sim、NaVILA checkpoint and inference server、benchmark scenes and robot assets、CUDA-capable workstation；硬件/传感器：NVIDIA GPU compatible with Isaac Sim、no physical robot required for benchmark evaluation；实机证据：This artifact evaluates simulation navigation and does not independently reproduce Go2 or Booster T1 hardware results.；开放范围/限制：Official co-author extension and useful benchmark path, but not an independent community reproduction and not physical-robot evidence. Repository MIT covers its code only; Isaac Sim, scene assets and upstream benchmark licenses must still be reviewed separately.
官方数据 · 许可证 No dataset card or artifact license disclosed; raw video/image rights remain with upstream sources · 最近核验 2026-07-22
依赖：VLN-CE and Habitat scene assets、R2R and RxR source annotations、yt-dlp for copyright-restricted Human Touring video IDs、NaVILA extraction and frame-render scripts、external VQA datasets following NVILA preparation；硬件/传感器：local storage for roughly 46.16GB of listed LFS objects、GPU rendering may be needed to reconstruct annotation-only VLN videos、the paper VLA training used multiple 8xA100 nodes；实机证据：Human Touring annotations improve reported Go2 and Booster T1 outdoor and multi-room navigation, but the raw copyrighted videos/images are not redistributed.、The repository explicitly says R2R augmented video is annotation-only and must be regenerated from upstream scenes.；开放范围/限制：The download portal is real, but it is not a self-contained training corpus: several raw media and simulator assets must be obtained from upstream sources, and the dataset repository has no license. Do not label it commercially reusable or fully reproducible.
官方权重 · 许可证 No model-card or repository license disclosed; upstream Llama/VILA terms also apply · 最近核验 2026-07-22
依赖：NaVILA fixed code stack、Python 3.10、PyTorch and CUDA、VILA/NVILA dependencies、Transformers 4.37.2、FlashAttention 2.5.8、WebDataset 0.1.103 for VLN-CE compatibility；硬件/传感器：single RTX 4090 for the documented roughly 1 FPS FP16 server、FP16 measurement 18.5GB GPU memory and 594.58ms for 1,737 context tokens plus 10 output tokens、W4A16 paper ablation reports 8.6GB and 367.80ms but no separately identified quantized checkpoint was found；实机证据：The authors deploy the same VLA on Unitree Go2 and Booster T1 across workspace, home and outdoor scenes; table cells report simple/complex instruction success rather than a single pooled headline.、An independent BSD-2-Clause StrikeRobot Go2 ROS2 workspace documents loading this checkpoint, eight-frame buffering, Blackwell fixes and a closed-loop motion controller, but it publishes no comparable navigation success-rate denominator.；开放范围/限制：Bytes and anonymous access were verified without downloading the 16.99GB repository. Missing artifact-specific license prevents treating the weights as unrestricted open source. The low-level locomotion stack and robot SDK integration remain platform-specific.

“官方”表示归属与论文关系已核验；仍应按所列依赖、硬件和许可证复现。

工程资源评审结论与反面证据

代码 · 已核验官方资源
审核码：verified_official; 官方关系：verified_official_repository; 最近核验：2026-07-20
代码：Repository and immutable commit 76b98f233dd0fff05dfcd69435eec6740febff9d were retrieved and matched to the paper.
权重：Not separately promoted by this code-repository review.
数据：Not separately promoted by this code-repository review.
许可证：Apache-2.0；Verified at commit 76b98f233dd0fff05dfcd69435eec6740febff9d. Model weights and datasets must be licensed and verified separately; Apache-2.0 is not automatically inherited by them.
依赖：Python 3.10、VILA training stack、FlashAttention 2、Transformers 4.37.2、VLN-CE and Habitat evaluation dependencies、WebDataset 0.1.103
硬件/传感器：One or more CUDA GPUs for the documented evaluation path、Legged robot hardware for physical navigation; exact SDK boundary is not fully packaged in the README
实机证据：The paper/project reports physical legged-robot navigation; the repository also links weights, datasets and an Isaac Sim benchmark, but this batch promotes code only.
独立复现：No independent runtime reproduction with fixed hardware and quantitative metrics was performed or established in this batch.
限制/反面证据：Verified at commit 76b98f233dd0fff05dfcd69435eec6740febff9d. Model weights and datasets must be licensed and verified separately; Apache-2.0 is not automatically inherited by them.
当前建议：工程资源已核验；Use the pinned repository as an engineering starting point, while preserving the stated license, dependency and evidence boundaries.
重点标签：artifact:code-verified、license:explicit、runtime:not-reproduced
核验来源：来源 1 · 来源 2

暂缓或驳回结论不会升级代码、权重或数据开放状态；Stars、转发和演示视频不等于独立复现。

关系纠错与反面证据

已移除误标来源条目 · 已纠错链接
旧里程碑标签没有“论文”字样，历史构建把RSS DOI误推断为项目页。该URL是正式论文出版标识；publisher与paper关系已由元数据终审保留。；最近核验 2026-07-22

纠错前关系仅作为审计证据保留；精确纠错不会删除同一 URL 上仍然成立的其他关系。

原始来源

最近核验：2026-07-22。链接存在不等于官方仓库、可运行、已开放权重/数据或许可证允许使用。