Joint Field Characterisation of Onboard Perception, Mapping and Manipulation on a Sub-13 kg Rocker-Bogie Rover

This paper reports the joint outdoor characterisation of detection, mapping, local re-planning and grasping running concurrently on one 4 GB embedded graphics processing unit module aboard a 12.982 kg six-wheel rocker-bogie rover, and the power it draws. Platforms of this class are characterised one subsystem at a time. The high-level stack runs as Robot Operating System 2 (ROS 2) nodes. On a held-out split of 250 annotated outdoor frames spanning overcast and direct sunlight, a half-precision single shot multibox detector on a MobileNetV2 backbone reached 89.4% mean average precision at 0.5 intersection over union while the vehicle sustained 14.5 frames per second with mapping, planning and telemetry resident. Planar light detection and ranging inertial mapping over a bounded 10 m by 10 m grass plot returned 11.2 cm positional root mean square error, with 220 ms local re-planning latency, bounding safe ground speed near 0.34 m s −1. Thirty autonomous pick-and-place trials on a 12 cm cylindrical target succeeded 23 times: 76.7%, Wilson 95% interval 0.59–0.88. Mean terminal draw under mixed duty was 38.5 W, exceeding the duty-weighted prediction from three isolated states by 16.3 W and giving about 5.3 h of derived endurance. Perception , mapping, manipulation and terminal power were measured together on one sub-13 kg field platform, and the shortfall shows that a budget built from isolated subsystem states underestimates what this autonomy actually draws.