ActiveFly-Bench Mixed Real/Sim Aerial VLA Dataset
真机室内、真机室外与EmbodiedCity仿真统一的空中EQA→观察规划→细粒度控制数据。固定Hugging Face修订公开非gated，143个文件约244.63GB；已下载的EQA/OBP JSON分别为3,919/982条。论文与当前发布的计数、真实test切分存在差异，聚合CC BY 4.0也不能覆盖所有上游模型、代码和仿真资产条款。
- 开放状态
- 项目制品开放·上游条款另适用
- 许可证
- CC BY 4.0 dataset-card declaration; upstream OpenVLA/OpenPi, simulator, environment assets and nested code licenses remain applicable
- 商业使用
- original released corpus appears usable with attribution, subject to separate upstream asset/model/code terms and aviation/safety regulation
- 人工评审结论
- verified_public_nongated_fixed_revision_244628442267_bytes_split_and_license_caveats
- 机器人
- custom 280 mm physical quadrotor、EmbodiedCity simulated UAV
- 任务
- Aerial Embodied Question Answering、Observation Behavior Planning、Fine-grained Language-guided UAV Control、active visual search、viewpoint adjustment
- 模态
- egocentric RGB frames/video、language instruction、multiple-choice EQA、observation-behavior label、UAV odometry x/y/z/yaw、gimbal state/action、trajectory waypoints、environment/source metadata
- 格式
- tar.gz trajectory archives、ZIP simulator assets、JSON EQA/OBP、Safetensors OpenVLA、Orbax OpenPi
- 采集/生成方式
- five professional pilots collect expert ROSbag trajectories; adaptive keypoints are perturbed with Gaussian position/yaw noise and replayed by Ego Planner/PID/gimbal control to generate 5-40 demonstrations per instruction; simulation uses matched task categories
- 已知问题
- Paper says about 1.3k EQA/1.3k OBP but downloaded fixed files contain 3,919/982 rows; paper says 80/20 across categories but current real archives are train-only while test is simulation-only; full 244.63GB not downloaded; raw LiDAR/IMU, calibration and failure balance unverified; repository lacks root LICENSE although README declares CC BY 4.0; no quantified physical success or independent reproduction.