UAV-Flow
UAV-Flow is the first real-world benchmark for language-conditioned, fine-grained UAV trajectory control, framed as the "Flying-on-a-Word" (Flow) task.
Source notes
UAV-Flow is the first real-world benchmark for language-conditioned, fine-grained UAV trajectory control, framed as the "Flying-on-a-Word" (Flow) task. Certified pilots — each with 800+ hours of experience — manually flew 30,692 short-range trajectories across three university campuses (5.02 km²) on DJI Mavic 3T RTK platforms, flying exclusively from the drone's first-person view. Every trajectory pairs an atomic natural-language instruction ("go through the lamp post from the right side") with egocentric onboard video and RTK-accurate 6-DoF state. Accepted at NeurIPS 2025; a companion simulated split is catalogued separately as UAV-Flow-Sim.
Note: the code is Apache-2.0 but neither HuggingFace dataset repo declares a data license. Total recording hours are not stated in the paper.
- Scale
- 258 GB size
- Formats
- custom
- License
- unknown
- Published
- 2025-05-14
Decision summary
teleoperation
Metadata requires review against the official source before publication.
Read the dataset manifest and feature schema.
Release history
Source-backed release timing for this canonical dataset record.
- Release evidence
Dataset series published
Release timing is recorded from official dataset metadata.
Selection readiness
6 evidence dimensions for deciding whether this dataset is ready to inspect, compare, or adopt. This is not a model benchmark.
Catalog evidence · not task performance
How scores work73/100
Provisional
4/6 dimensions scored · 48% confidence
Read the evidence behind all 6 dimensions
Access and governance
Access and license are declared by the source.
Schema and signal coverage
No machine-readable schema has been verified yet.
Policy training readiness
Observation and action/state signals are declared; alignment quality still depends on sample verification.
World-model readiness
Temporal observations plus geometry or semantic context are declared; sample alignment remains to be audited.
Failure and recovery readiness
No verified failure/recovery annotation evidence is available yet.
Download and processing readiness
Scale is declared; transfer and processing estimates are not measured.
Review unresolved evidence and next checks
Signal gaps
- Gaze / attention. Not enough evidence is available to classify this signal.
- Feedback / correction / failure. Not enough evidence is available to classify this signal.
Next checks
- Read the dataset manifest and feature schema.
- Run a bounded sample audit before assigning Strong readiness.
Engineering checks (OBRS) behind this record
Needs Audit
OBRS metadata and reported test evidence. A breakdown behind Selection readiness, not a separate score or an independent OpenBot certification.
- A passing real-hardware test report is required.
- A passing ingestion/pipeline test report is required.
- A passing privacy and provenance review is required.
Dataset facts
- Source
- Beihang University — Institute of Artificial Intelligence
- Evidence
- secondary claim
- Formats
- custom
- episodes
- 30.7K
- tasks
- 8
- size
- 258 GB
Loop signals
5/7 present or partial
No decision-grade evidence captured yet.
unknownNo decision-grade evidence captured yet.
unknownView 5 more signal categories
video · UAV-Flow is the first real-world benchmark for language-conditioned, fine-grained UAV trajectory control, framed as the "Flying-on-a-Word" (Flow) task. Certified pilots — each with 800+ hours of experience — manually flew 30,692 short-range trajectories across three university campuses (5.02 km²) on DJI Mavic 3T RTK platforms, flying exclusively from the drone's first-person view. Every trajectory pairs an atomic natural-language instruction ("go through the lamp post from the right side") with egocentric onboard video and RTK-accurate 6-DoF state. Accepted at NeurIPS 2025; a companion simulated split is catalogued separately as UAV-Flow-Sim. *Note: the code is Apache-2.0 but neither HuggingFace dataset repo declares a data license. Total recording hours are not stated in the paper.*
presentUAV-Flow is the first real-world benchmark for language-conditioned, fine-grained UAV trajectory control, framed as the "Flying-on-a-Word" (Flow) task. Certified pilots — each with 800+ hours of experience — manually flew 30,692 short-range trajectories across three university campuses (5.02 km²) on DJI Mavic 3T RTK platforms, flying exclusively from the drone's first-person view. Every trajectory pairs an atomic natural-language instruction ("go through the lamp post from the right side") with egocentric onboard video and RTK-accurate 6-DoF state. Accepted at NeurIPS 2025; a companion simulated split is catalogued separately as UAV-Flow-Sim. *Note: the code is Apache-2.0 but neither HuggingFace dataset repo declares a data license. Total recording hours are not stated in the paper.*
presentlanguage · UAV-Flow is the first real-world benchmark for language-conditioned, fine-grained UAV trajectory control, framed as the "Flying-on-a-Word" (Flow) task. Certified pilots — each with 800+ hours of experience — manually flew 30,692 short-range trajectories across three university campuses (5.02 km²) on DJI Mavic 3T RTK platforms, flying exclusively from the drone's first-person view. Every trajectory pairs an atomic natural-language instruction ("go through the lamp post from the right side") with egocentric onboard video and RTK-accurate 6-DoF state. Accepted at NeurIPS 2025; a companion simulated split is catalogued separately as UAV-Flow-Sim. *Note: the code is Apache-2.0 but neither HuggingFace dataset repo declares a data license. Total recording hours are not stated in the paper.*
presentUAV-Flow is the first real-world benchmark for language-conditioned, fine-grained UAV trajectory control, framed as the "Flying-on-a-Word" (Flow) task. Certified pilots — each with 800+ hours of experience — manually flew 30,692 short-range trajectories across three university campuses (5.02 km²) on DJI Mavic 3T RTK platforms, flying exclusively from the drone's first-person view. Every trajectory pairs an atomic natural-language instruction ("go through the lamp post from the right side") with egocentric onboard video and RTK-accurate 6-DoF state. Accepted at NeurIPS 2025; a companion simulated split is catalogued separately as UAV-Flow-Sim. *Note: the code is Apache-2.0 but neither HuggingFace dataset repo declares a data license. Total recording hours are not stated in the paper.*
presentOpen · unknown · custom
presentEvidence details and provenance
Official signal claims
Schema and annotations
No machine-readable schema facts are captured.
Sample verification
Pending. Metadata does not prove sample coverage, alignment, or file integrity.
curated official source
Integration notes
- Metadata requires review against the official source before publication.