A Lightweight Vision-Language-Action Policy with Progress-Aware Hybrid Execution for UAV Waypoint Navigation in...

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Offline action prediction in vision-language-action (VLA) models does not guarantee reliable closed-loop flight for unmanned aerial vehicles (UAVs).
• Existing VLA models struggle with real-time control and task completion in dynamic environments.

Purpose of the Study:

• To evaluate the effectiveness of a progress-aware hybrid execution system for UAV closed-loop control.
• To investigate the limitations of raw VLA policy control in a simulated waypoint task.
• To assess the impact of checkpoint sensitivity and system integration on overall performance.

Main Methods:

• Utilized the AirSim Blocks waypoint task with 100 expert episodes and 3385 RGB-D, instruction, state, and action records.
• Trained a 132,840-parameter VLA policy and evaluated its performance on seen and unseen validation data.
• Embedded the VLA policy's action proposals within a hybrid execution framework incorporating explicit recovery and near-goal precision mechanisms.

Main Results:

• The trained VLA policy achieved high action accuracy (0.9278±0.0019) but failed in raw closed-loop control.
• The hybrid system demonstrated significant improvement, with the strongest checkpoint achieving 59/60 goals.
• A conservative estimate across multiple checkpoints and seeds yielded 147/180 successes (81.7%) with zero collisions, highlighting system robustness and checkpoint sensitivity.

Conclusions:

• Closed-loop control for UAVs necessitates sophisticated hybrid systems that go beyond raw VLA policy outputs.
• Progress-aware execution and explicit recovery strategies are vital for reliable task completion and collision avoidance.
• The performance of the VLA policy is heavily influenced by the integration within the broader hybrid execution system, not solely by its learned policy.