To operate independently, each platform integrates an onboard computation unit, a two-degree-of-freedom motorized gimbal with active cooling, and a battery pack.
To power the mobile platform, we design a compact battery module using a single 18650 lithium-ion cell.
We use a Raspberry Pi 4B as the onboard compute platform, providing access to communication protocols including I2C, SPI, UART, PWM, GPIO, and high-speed USB devices.
The top plate embeds an active radiator to prevent thermal throttling when the Raspberry Pi operates at high power.
The gimbal expands the field-of-view of onboard sensors by providing pitch and roll adjustments.
We use a charging dock with magnetic pogo pins, allowing mobile robots to recharge these modules on demand.
We use a suction-based base to deploy the module to a wide variety of surfaces. The module connects to the suction mount through a magnetic ring integrated into the back of the module.
The SensorPerch platform weighs
201.4g
external dimensions
99mm×66mm×67mm
The total hardware cost is approximately
$94*
- Raspberry Pi 4B$60
- two MG90S servos$10
- a blower fan$4
- a PCB$5
- a 18650 battery cell$6
- a pogo-pin connector$4
- miscellaneous hardware and filament$5
* RealSense D435i not included
Our onboard battery supports
2.25hours
of continuous streaming with the RealSense D435i
shear resistance against package weight
12×
weight resistance in pull-off
4×
Given the total platform weight of 298.7 g with the RealSense D435i payload
average streaming latency
180ms
With a RealSense D435i camera
SensorPerch reaches ~80% success across all three, closely matching the oracle viewpoint and confirming that the autonomously constructed views align with the policy's training distribution.
Across 10 trials, we reliably detect task-relevant object states including boiling water and an empty cup (90% success rate) and successfully support cutting policy execution (80% success rate).
SensorPerch treats robot perception as a problem of physically constructing task-relevant viewpoints, rather than constraining sensing to the robot’s embodiment or fixed infrastructure. By modeling sensors as reconfigurable entities, we enable perception to adapt to task demands.
BibTeX
@misc{wu2026sensorperchsensematters,
title={SensorPerch: Sense Wherever and Whenever it Matters},
author={Zhanxin Wu and Ruofei Tong and Tapomayukh Bhattacharjee},
year={2026},
eprint={2607.10682},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2607.10682}
}