A drone learns to carry out short flight commands
UAV-Flow++: A Unified Suite for Benchmarking and Learning Language-Guided UAV Flight · arXiv:2611.17163
In plain words
A drone may understand a destination yet still need to work out how to move around nearby objects to carry out a short command. UAV-Flow++ pairs expert flight examples with a simulator and a controller called FlowVLA, which uses instructions, images, and the drone’s state to choose actions.
In a simulated test, further training helped FlowVLA complete more instructions fully. The test used one campus scene, so it does not show how well the controller handles unfamiliar scenes.
The research
A drone that can identify a destination may still need to interpret motion and spatial references to move safely nearby.
This work tests whether a controller can carry out those short, instructed maneuvers, and whether training through simulated flight improves its chance of meeting the whole instruction.
UAV-Flow++ pairs real flight demonstrations with simulation and a learned controller to test how drones turn short language instructions into nearby flight maneuvers.
UAV-Flow-Sim replays each instruction from a fixed starting state. Its automated score checks both the maneuver and required final state; on 273 episodes, FlowVLA averaged 75.58% success against 68.86% for adapted π₀-UAV.
That comparison changes architecture and training data together, so it cannot isolate either cause. Starting from the same FlowVLA, GRPO and TGRPO post-training reached 92.06% and 93.16%, showing higher overall task completion.
The pretrained-versus-baseline comparison changes both architecture and training-data mix; the post-training results come from simulation, whose test episodes share the simulator’s campus scene.
What the evidence shows
- From destination to maneuver
UAV-Flow++ is a benchmark and learning suite for short-range, language-guided UAV maneuvers. It tests whether a drone can interpret a concise instruction, relate it to what it sees, and execute feasible flight actions.
This targets the gap between choosing a destination and carrying out the nearby movement needed to reach or interact with it.
- A test of complete actions
The suite combines expert-piloted flight examples with an interactive simulator and a trained controller, FlowVLA. In the simulator, each action affects what the drone sees next, so evaluation reflects a sequence of actions rather than a single prediction.
An automated evaluator counts an episode as successful only if the maneuver and final state satisfy all applicable task criteria.
- The baseline comparison
On the simulated test, FlowVLA had a larger share of episodes meeting every task-completion criterion than adapted π₀-UAV. Because their architectures and training data differed, the comparison cannot identify which difference explains the advantage.
- What post-training adds
The post-training comparison starts GRPO and TGRPO from the same pretrained FlowVLA. Their mean success rates were 92.06% and 93.16%, respectively, on the full simulated test set.
This supports the conclusion that post-training improved overall completion, with a smaller additional success-rate gain for TGRPO over GRPO.
- Where the result stops
The result is an aggregate over the full test set, while reinforcement-learning training covered Pass, Rotate, Land, and Approach. It does not imply equal gains on every task.
The test episodes came from the same campus scene used for simulated demonstrations, and the paper identifies reliance on ground-station computation and wireless communication as a practical constraint on physical flight.
Deep
- Original paper authors
- Donglin Yang, Xiangyu Wang, Yantong Zhong, Wenhao Zheng, Zihan Ding
- Submitted to arXiv
- License
- CC BY 4.0