Study a defined model.
Useful for controlled questions. Its assumptions constrain how the result can transfer.
Research
Reliable vision
starts with evidence.
We study how an onboard system acquires a designated target, maintains a useful track, recognises loss and recovers. Each result belongs to its test conditions.
Research notes
How we work.
What we learn.
Notes on evaluation and engineering practice, with the scope of the work stated directly.
14 September 2026
How we separate development results from independent evaluation.
Keeping data roles, model decisions and evidence categories distinct throughout development.
Research practice / Version 1.0Recorded development
A view into
the research bench.
This player pairs external source footage with saved visual observations from an existing embedded-compute development replay.
Recorded footage. Recorded output.
Inspect the observation.
The overlay comes from an existing Jetson replay using a borrowed detector baseline and a software-generated cue. Play or scrub the video to see the saved observations.
Recorded overlay / Ready
This sample illustrates the replay format. The recorded imagery does not respond to guidance, and it does not establish flight readiness, real-radar performance or interception effectiveness.
Source and evaluation context
The complete development clip contains 310 frames at 30 fps, retained at its original 640 × 512 geometry. The player displays saved boxes; no model runs in your browser. Missing visual observations remain blank.
Our wider evaluation has exposed failures across domains and competing objects. No integrated autonomous project-drone flight is claimed.
The overlay appears in the inline player. Native full-screen video shows the original footage.
Evidence categories
Different tests.
Different conclusions.
Inspect repeatable inputs.
The same sequence can be replayed and compared. It cannot establish how a physical system responds in flight.
Measure the actual hardware.
Real cameras and computing expose practical behaviour. Integrated flight is a separate evidence step.
Protocol in preparation
Towards
Cue-to-Track.
A proposed evaluation of cue-conditioned aerial tracking on edge hardware: acquisition, track continuity, loss, reacquisition and embedded timing.
The protocol is in preparation. It is not a released benchmark or peer-reviewed paper. Its intended scope stops before closed-loop flight and interception claims.
Discuss research collaboration