Roadmap¶

Proven now¶

Next engineering milestones¶

Close the performance gap¶

Sweep 320, 256 and 192 pixel inputs against 3D error
Test state-guided ROI cropping across 12 motion profiles; fixed 320 × 320 export showed no inference saving
Test frame skipping; throughput improved but fast/nonlinear motion error was unacceptable
Export a dynamic-shape or smaller-input model that can make ROI inference cheaper
Add accuracy-gated, motion-aware inference cadence and detector scheduling
Explore double buffering around the single NPU queue
Improve NPU memory configuration and DMA/zero-copy data paths
Use quantization-aware training to recover INT8 accuracy

Broaden evidence¶

Re-run calibration refinement with the explicit detector offset and validate the proposed extrinsics end to end
Record varied flight trajectories rather than hover only
Separate quantization holdout from the final accuracy corpus
Run 10–15 minute sustained thermal tests
Measure live network-to-state latency and jitter
Validate against real USB or IP cameras

Move toward deployment¶

Define explicit loss-of-tracking and landing safety policy
Calibrate a physical multi-camera room
Harden board networking and service supervision
Evaluate multi-room handoff and multi-target identity

The current honest status¶

The classical detector is fast and precise on i.MX93, but it needs a marker. The final submitted learned NPU path is fully delegated, reaches 6.25 fixes/s and reduces the complete-loop latency by 6.11× versus the two-core FP32 baseline. The next milestone is not another isolated inference number; it is a learned path that improves speed while recovering quantized accuracy across representative motion and supports closed-loop point-to-point flight.