Embodied VLN 200-Episode Real-World Evaluation Set (not released)

论文在10个真实场景执行200回合VLN横评：分层Insta360+LiDAR+SLAM配置为51% SR/37%严格语义SR/7%碰撞率，单体D435i RGB配置为22%/17%/51%。原始指令、图像、LiDAR、轨迹、碰撞和语义终止标签均未发布。

开放状态: 项目已核验·制品未开放
许可证: not available for the absent evaluation package
商业使用: blocked
人工评审结论: paper_verified_200_physical_navigation_episodes_not_released
机器人: custom differential-drive VLN robot
任务: step-by-step vision-language navigation、intent-grounded navigation、backtracking、semantic stopping、collision avoidance
模态: Insta360 X4 panorama、RealSense D435i RGB-D camera used as RGB for monolithic input、Livox Mid-360 LiDAR、language instruction、SLAM pose、navigation action、collision and semantic-stop outcome described
格式: not released
采集/生成方式: physical closed-loop navigation with onboard Jetson Orin Nano and remote NVIDIA L40 inference over Wi-Fi
已知问题: No raw files, manifest, checksum, privacy treatment or dataset license; one run per episode; fixed seed 42; system configurations are not sensor- or controller-matched; no independent reproduction.