论文记录 · 证据边界公开

ActiveFly-Bench: Aligning Embodied Question Answering with Vision-Language-Action for Aerial Embodied Perception

arXiv technical report · 2026-07-11 · 通用或轻量 VLA

论文元数据

作者: Weichen Zhang、Shiquan Yu、Yinan Zhu、Peizhi Tang、Shilong Ji、Zhiyuan Deng、Tianyi Lyu、Haoyang Wang、Xin Zeng、Chen Gao、Yong Li、Xinlei Chen
机构: 机构尚未补齐
发布日期精度: 日
研究方向: 通用或轻量 VLA
机器人: EmbodiedCity simulated UAV、custom 280 mm physical quadrotor
DOI: 未披露
arXiv: 2607.10180
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: 存在链接线索，尚未核验官方性与可运行性
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: 许可证条款已核验（不等于可商用）

当前实践优先级

高潜待持续验证 · 73/100 · 0–6 个月

路线标签：近期热门、可能成为新主流、导航VLA技术2.0·跨域主动感知分支·代码数据权重开放·量化真机仍缺

社区证据类型：尚无独立社区验证

ActiveFly把EQA、观察规划和细粒度控制统一成可运行流水线，真实发布约245GB混合数据、两类权重和ROS客户端；但真实飞行只有案例与390ms时延，论文/发布计数和test切分不完全一致。

近期窗口 · 35/35
发布距评分基准日 0 个月，落入 0–6 个月窗口。
社区可复查验证 · 4/25
固定代码仓、公开非gated数据/权重和样本JSON可查；尚无独立训练、仿真或实机复现。
跨领域技术创新 · 20/20
把空中具身问答、观察行为推理、VLA细粒度控制、SLAM/规划和真机网络闭环连接成跨域主动感知路线。
量化实机证据 · 4/10
给出自制四旋翼硬件和5Mbps下390ms环时延，但没有实机次数、成功率、碰撞/急停或失联恢复。
开放与可复现 · 5/5
训练/评测/ROS客户端、两类权重和约245GB数据真实可下；飞控/规划缺失，根许可证、切分和计数有边界。
工程实用增量 · 5/5
分层模块便于替换VLM/VLA和接入经典安全导航栈；当前2.56Hz stop-and-infer与Oracle Success口径限制直接部署。

社区可复查证据：尚未收录；Stars、forks 与作者演示不计社区验证。

证据缺口：缺少显式可复查社区复现/部署证据；量化实机证据不足；真实飞行无trial denominator、SR、碰撞、失联或安全指标；论文1.3k/1.3k EQA/OBP与发布3,919/982不一致；当前真实数据只有train归档而test只有仿真；README称CC BY 4.0但根LICENSE缺失且上游条款分层；飞控、路径规划、完整标定和独立复现缺失

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 README declares CC BY 4.0 but no root LICENSE file exists; ActiveFly-Client-ROS is BSD-3-Clause, the vendored OpenVLA run tree is MIT, and OpenPi/other upstream components retain separate terms · 最近核验 2026-07-22
依赖：Python 3.11、PyTorch、OpenVLA or OpenPi stacks、flash-attn 2.5.5 for OpenVLA、CUDA GPU、Hugging Face dataset archives、commercial or local VLM endpoint for Air-EQA/OBP、ROS for physical VLA client；硬件/传感器：CUDA GPU suitable for OpenVLA/OpenPi; exact minimum VRAM not disclosed、about 245GB storage for full artifact host、for author physical setup: 280mm quadrotor, Intel NUC13, PX4, Livox Mid-360 and gimballed camera、flight-control and path-planning modules must be supplied separately；实机证据：The paper shows author indoor/outdoor closed-loop UAV cases and about 390ms control-loop latency at 5Mbps; there is no physical trial denominator, success rate, collision rate or independent flight reproduction.；开放范围/限制：Use the fixed repository for benchmark training/evaluation and simulation. Treat real-flight code as a sample VLA client only, audit every nested/upstream license, and do not operate a physical UAV without an independently verified flight controller, geofence, loss-link behavior and safety case.
官方数据集 · 许可证 CC BY 4.0 dataset-card declaration; simulator assets, upstream environments, model outputs and code retain separate terms · 最近核验 2026-07-22
依赖：Hugging Face Hub or Xet download、about 245GB storage for the full host revision、tar/zip extraction、ActiveFly evaluation code、EmbodiedCity simulator for simulation rollout、upstream license audit；硬件/传感器：large local or object storage、CPU/RAM for large archive extraction、CUDA GPU for VLA training/evaluation、physical UAV not required for offline dataset inspection；实机证据：The corpus includes author-collected real indoor/outdoor trajectories from a custom quadrotor; archive-internal sensor calibration and failure balance were not fully downloaded/audited.；开放范围/限制：Use as one canonical mixed dataset visible on both real and simulation pages. Preserve the paper-versus-release count and split mismatch, do not claim that all 245GB were downloaded, and do not assume raw LiDAR/IMU or failed trajectories exist until archive sampling proves them.
官方权重 · 许可证 Aggregate dataset card declares CC BY 4.0; OpenVLA/OpenPi base-model and derivative terms remain applicable and no separate checkpoint model card is provided · 最近核验 2026-07-22
依赖：OpenVLA or OpenPi code from the fixed repository、PyTorch/JAX as required by the selected checkpoint、CUDA accelerator、safetensors or Orbax、matching normalization statistics and ActiveFly data paths；硬件/传感器：about 43.34GB checkpoint payload for both families、RTX A6000 used for author real deployment、minimum supported VRAM, batching and end-to-end latency per checkpoint are not disclosed；实机证据：Author physical deployment uses OpenVLA on an RTX A6000 ground station with stop-and-infer; the released weights are not a complete autonomous flight package.；开放范围/限制：Pin the host revision and selected model family, audit base-model terms, and measure terminal SR plus safety events rather than relying on oracle success. Do not infer that the aggregate dataset license overrides checkpoint lineage.

“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。

每日候选人工终审

ActiveFly-Bench把空中具身感知拆成Air-EQA、观察行为规划和细粒度语言控制三层，并真实发布仿真/真机轨迹、OpenVLA/OpenPi权重、训练评测代码及ROS客户端。论文图3给出10,076条FLUC，其中5,945条仿真、2,228条真机室内、1,903条真机室外；表2显示GPT-5.4+Pi0.5端到端EP为47.8，GPT-5.4+OpenVLA为18.9，人类+Pi0.5为60.3，而全人类为97.3。Pi0.5在FLUC的SR/OSR/nDTW/NE为31.0/71.0/12.9/9.86，说明到过目标区域并不等于终态控制成功。作者在自制280毫米四旋翼上展示室内外闭环，但只给约390毫秒单控制环时延，没有真实任务重复数、成功率、碰撞率或失联恢复率；视频/状态声称30 Hz传输与stop-and-infer约2.56 Hz决策必须区分。固定代码仓含249个blob和训练、评测、ROS客户端，但不含飞控与路径规划；README称CC BY 4.0而根目录无LICENSE，OpenVLA/OpenPi及两个嵌套许可证仍需逐项遵守。固定Hugging Face修订公开非gated，143个文件约244.63 GB；下载样本确认EQA为3,919条、OBP为982条，与论文约1.3k/1.3k口径不一致，且当前清单中的真实轨迹只有train归档、test只见仿真。因此它是近期跨域导航/VLA高价值基础设施，但真实性、切分、许可证和真机安全口径仍需显式展示。

实机证据：author_qualitative_closed_loop_physical_uav_with_latency_breakdown_but_no_trial_denominator_or_success_rate；作者用GPT-5.4做问答、OpenVLA预测动作，在室内外展示闭环空中具身感知。机载FAST-LIO2与Ego Planner负责定位和避障，VLM/VLA在A6000地面站运行；5 Mbps Wi-Fi下图像传输约120 ms、VLA约250 ms、规划/PID各约10 ms，总计约390 ms。论文没有实机episode数、SR、碰撞/急停、失联、风扰或电量分母，不能据演示视频推断稳健部署。

硬件/本体：custom 280 mm carbon-fiber quadrotor、T-Motor F90、Intel NUC 13 ANKB、PX4、Livox Mid-360、XF C-200T three-axis gimbal、RTX A6000 ground server

红旗与边界：真机只有案例图和390 ms时延，没有episode数、成功率、碰撞率、失联恢复或置信区间；论文的30 Hz视频/状态传输不能等同控制决策频率；stop-and-infer约390 ms即约2.56 Hz；EP中的FLUC成功采用Oracle Success，不是终态SR，容易高估闭环控制；论文约1.3k Air-EQA/1.3k OBP与固定修订3,919/982条不一致；论文称各类80/20切分，但固定修订当前只有仿真test归档，真实数据只见train；代码仓README称CC BY 4.0但根LICENSE缺失，嵌套OpenVLA/OpenPi/ROS代码条款不同；真实部署代码只含VLA客户端，不含飞控、FAST-LIO2/Ego Planner集成和完整安全闭环；245 GB未全量下载，归档内帧率、标定、LiDAR/IMU原始数据和失败轨迹尚未逐文件审计；没有独立团队复现

查看资源 · code · 状态 official_training_evaluation_simulation_and_vla_client_source_fixed_commit_with_mixed_license_boundary · 许可证 README declares CC BY 4.0 but no root LICENSE file exists; ActiveFly-Client-ROS has BSD-3-Clause and vendored OpenVLA has MIT; OpenPi and other upstream components retain their own terms
固定提交498f426含329路径、249个blob、约2,591,882字节，覆盖Air-EQA/OBP评测、OpenVLA/OpenPi训练与rollout、指标、流水线UI和ROS VLA/gimbal客户端。README明确不含飞控与路径规划；无CI、release或完整端到端安全测试。
查看资源 · weights · 状态 official_openvla_and_openpi_weights_public_nongated_fixed_revision · 许可证 Hugging Face dataset card declares CC BY 4.0; OpenVLA/OpenPi base-model and derivative terms remain applicable and are not replaced by the aggregate card
固定修订weights目录约43,335,727,342字节，含OpenVLA四个safetensors分片及OpenPi Orbax参数/训练状态。没有独立模型卡、参数量/显存/机器人安全适配说明。
查看资源 · data · 状态 official_real_and_sim_trajectory_archives_public_nongated_fixed_revision_with_split_and_count_mismatch · 许可证 CC BY 4.0 dataset-card declaration; simulator assets, source environments and third-party model outputs require upstream-rights review
固定修订总计143个文件、244,628,442,267字节，其中datasets约184.11 GB、Sim_Envs约17.19 GB、weights约43.34 GB。已实际下载并解析1.92 MB EQA.json与0.82 MB OBP.json，分别为3,919和982条；真实轨迹当前只见train归档，test只见仿真。完整245 GB未下载。

论文入库不等于代码、权重或数据开放；“承诺开放”“声称开放”和“已核验可下载”严格分开。

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。