Project: #411 Bridge Avoidance in River-based Drone Autonomy Progress Report - Reporting Period Ending: March 30, 2023 Principal Investigator: Mahadev Satyanarayanan Status: Active Start Date: Dec. 1, 2022 End Date: June 30, 2023 Research Type: Advanced Grant Type: Research Grant Program: FAST Act - Mobility National (2016 - 2022) Grant Cycle: 2022 Mobility21 UTC Progress Report (Last Updated: March 27, 2023, 11:33 a.m.) % Project Completed to Date: 30 % Grant Award Expended: 22 % Match Expended & Document: 100 USDOT Requirements Accomplishments Our research to date has focused on measuring the limitations of edge offload from the drone to the cloudlet. This will determine the maximum complexity of the obstacle detection algorithms that can be executed in real time on the drone. Video from the drone consists of a 720p UDP RTSP stream at 30 FPS. Although RTSP is an Internet standard, the drone does not use a standard keyframe-based stream encoding scheme. Instead, the vendor's proprietary encoding distributes slices of each video frame across a number of network packets. Since UDP RTSP does not perform packet retransmission, this reduces the visual impact of packet loss. A negative consequence of this encoding is that software decoding has to be performed before individual frames can be selectively dropped. Our experiments show that LTE wireless transmission is a significant thermal bottleneck in the processing pipeline. Just receiving the RTSP stream over WiFi and retransmitting it via LTE causes thermal shutdown in 68 seconds. A transmission rate of 0.7 FPS stream is the maximum feasible without suffering thermal shutdown. This defines the upper bound of transmission rate available to the algorithms we use for bridge avoidance. Our results to date show that the extreme austerity and capacity of on-board processing and wireless communication are first-order design considerations. Weight and thermal issues constrain both onboard processing and offloading. Although the ground-based cloudlet can be very powerful, only modest use of its resources can be made while the drone is in flight --- at most one 720p frame's worth of image processing per second. This limits the agility of the drone in active vision settings. The next phase of our research will focus on the design of obstacle avoidance algorithms that are feasible even in the face of this extreme austerity. Impacts Our work is still in the early stages, and therefore much too soon to have impact. Other N/A Outcomes New Partners N/A Issues N/A