Tutorial

UAV Configurations:

Autonomy Algorithm: MobileNet

UAV Type: Asctec Pelican

Onboard Computer: Nvidia TX2

Sensor: PrimeSense RGB-D Camera (20 FPS with d=4m)

Analysis:

Performance is bottlenecked by sensor throughput. It adds a new ceiling.

Optimization Strategy:

Increase the sensor FPS.

UAV Configurations:

Increase sensor throughput to 33 FPS either by selection or optimization.

Analysis:

Performance is bottlenecked by UAV physics. i.e., the UAV doesn't have a higher acceleration to move any faster. Reducing the payload weight (heatsink weight) can help.

Optimization Strategy:

As a computer architect, reducing the TDP of the compute will lower the heat sink weight thus reducing the overall payload weight.

UAV Configurations:

Reduce the TDP by to 5 W by selecting a new onboard compute or by any optimization.

Analysis:

Performance is bottlenecked by sensors again. This could be another space to pursue further optimizations.

Optimization Strategy:

We need to be more aggressive in our optimization technique to reduce the TDP and increase sensor FPS further.

UAV Configurations:

Reduce the TDP to 1 W and increase runtime to 28 ms by selecting a new onboard compute or by any optimization. Increase sensor throughput to 40 FPS either by selection or optimization.

Analysis:

Performance is bottlenecked by UAV physics. i.e., the UAV doesn't have a higher acceleration to move any faster. However, the design point is near to the roofline kneepoint which represents the optimal design.

Optimization Strategy:

We need to be more aggressive in our optimization technique to reduce the TDP further.