Dual-ESP32 stereo for ARMOS TurtleBot: Design, synchronization, and ROS integration

We present a reproducible, low-cost stereo perception stack for the ARMOS TurtleBot that couples two OV7670 cameras to two ESP-WROOM-32 microcontrollers and a ROS pipeline on a Raspberry-Pi. A rigid T-mount fixes a 75 mm baseline; each sensor streams QVGA grayscale via an 8-bit DVP bus to its dedicated ESP32 using I2S-parallel with DMA. A shared pixel clock and microsecond VSYNC time-stamps enforce pairing with a 1 ms threshold, and frames are transported to the host using UDP (nominal) or UART (fallback). The method comprises photometric standardization, stereo calibration and rectification, SGBM disparity, depth recovery, and publication of costmaps for navigation; we report throughput, latency, synchronization, and depth accuracy across indoor scenarios. With QVGA over UDP, the pipeline sustains 12 fps at disparity, attains a median end-to-end latency near 94 ms (95th percentile under 120 ms), and yields costmap hit rates above 95% within 1.5 m. Mean absolute depth error grows with range (1.7 cm at 0.5 m to 12.8 cm at 3.0 m), consistent with the pinhole model. Ablations show that QQVGA over UART lowers median latency but reduces frequency and accuracy, while light JPEG reduces bandwidth variance at a modest long-range penalty. Mechanical drawings, bill of materials, firmware, and ROS configurations are provided to support replication.