- UAV-TRACKVLA // W7
- Python
- ANIMA
UAV-TRACKVLA
Language-Commanded Pursuit That Doesn't Lose Its Target
Persistent surveillance UAVs lose tracked targets during occlusions, direction changes, or handoffs between operators. Re-acquisition requires human attention and precious seconds. In a dynamic pursuit — a vehicle fleeing a checkpoint, a person of interest moving through a crowd — those seconds determine whether the mission succeeds. UAV-TRACKVLA (codename: Tengu) implements UAV-Track VLA (arXiv:2604.02241), fusing a frozen SigLIP visual backbone with temporal compression and a dual-decoder head: one for grounding (where is the target) and one for flow-matching action generation (how does the UAV move to stay on it).
MODULE STATUS: PRODUCTIONArchitecture
1 pass
- DIVISION
- GENERIC
- WAVE
- W7
- DOMAIN
- GENERAL
- WAVE 7 // ANIMA SUITE
- FOUNDATION — UAV-TRACKVLA
THE PROBLEM WE SOLVE
Persistent surveillance UAVs lose tracked targets during occlusions, direction changes, or handoffs between operators. Re-acquisition requires human attention and precious seconds. In a dynamic pursuit — a vehicle fleeing a checkpoint, a person of interest moving through a crowd — those seconds determine whether the mission succeeds.
A single operator can task a pursuit UAV in plain language and trust it to maintain lock through complex maneuvers — shrinking the human-machine ratio for ISR sorties and reducing cognitive load during high-tempo operations.
WHAT UAV-TRACKVLA DELIVERS
UAV-TRACKVLA (codename: Tengu) implements UAV-Track VLA (arXiv:2604.02241), fusing a frozen SigLIP visual backbone with temporal compression and a dual-decoder head: one for grounding (where is the target) and one for flow-matching action generation (how does the UAV move to stay on it).
CAPABILITIES
- UAV-TRACKVLA (codename: Tengu) implements UAV-Track VLA (arXiv:2604.02241), fusing a frozen SigLIP visual backbone with temporal compression and a dual-decoder head: one for grounding (where is the target) and one for flow-matching action generation (how does the UAV move to stay on it)
- Natural language commands update the grounding target in real-time, while the action decoder produces continuous flight commands without a separate planning module
- The unified architecture eliminates the latency of chained detect-plan-act pipelines.
WHY THIS IS HARD
Building UAV-TRACKVLA requires solving multiple coupled problems:
- 01Persistent surveillance UAVs lose tracked targets during occlusions, direction changes, or handoffs between operators
- 02Re-acquisition requires human attention and precious seconds
- 03UAV-TRACKVLA (codename: Tengu) implements UAV-Track VLA (arXiv:2604.02241), fusing a frozen SigLIP visual backbone with temporal compression and a dual-decoder head: one for grounding (where is the target) and one for flow-matching action generation (how does the UAV move to stay on it)
- 04Natural language commands update the grounding target in real-time, while the action decoder produces continuous flight commands without a separate planning module
UAV-TRACKVLA solves these through careful architecture design and rigorous validation.
PROOF, NOT PROMISES
Key performance metrics:
| METRIC | VALUE |
|---|---|
| Architecture | Unified VLA — grounding + action generation in one forward pass, no pipeline chaining |
| Action Generation | Flow-matching decoder produces smooth continuous flight commands — no discrete planning lag |
| Backbone | Frozen SigLIP (no fine-tuning cost) — only temporal + decoder heads are trained |
| Defense Angle | Persistent ISR pursuit, checkpoint escape re-acquisition, multi-operator UAV tasking |
WHAT'S BUILT TODAY
| COMPONENT | STATUS | NOTES |
|---|---|---|
| Architecture | COMPLETE | Unified VLA — grounding + action generation in one forward pass, no pipeline chaining |
| Action Generation | COMPLETE | Flow-matching decoder produces smooth continuous flight commands — no discrete planning lag |
| Backbone | IN PROGRESS | Frozen SigLIP (no fine-tuning cost) — only temporal + decoder heads are trained |
| Defense Angle | IN PROGRESS | Persistent ISR pursuit, checkpoint escape re-acquisition, multi-operator UAV tasking |
WHERE UAV-TRACKVLA DEPLOYS
- APP_01
AUTONOMOUS SYSTEMS
A single operator can task a pursuit UAV in plain language and trust it to maintain lock through complex maneuvers — shrinking the human-machine ratio for ISR sorties and reducing cognitive load during high-tempo operations.
- APP_02
RESEARCH LABS
UAV-TRACKVLA (codename: Tengu) implements UAV-Track VLA (arXiv:2604.02241), fusing a frozen SigLIP visual backbone with temporal compression and a dual-decoder head: one for grounding (where is the target) and one for flow-matching action generation (how does the UAV move to stay on it).
- APP_03
EDGE COMPUTING
Aerial tracker that understands "follow the white SUV turning left" and executes it — fusing language, vision, and flight control in one frozen backbone.