Sujal Bhakare
All projects

UB Lunabotics rover

Autonomous excavation rover for NASA's Lunabotics Challenge. Each subsystem runs on its own STM32 node over CAN, under a ten-package ROS2 autonomy stack. Best First Year Team, 2025.

Year
2024
Status
In progress
Role
Co-founder, Project Manager (previously Computer Hardware Lead)
Domain
Robotics, Embedded

Electronics

The rover is a set of independent modules. Each subsystem controls itself locally; only commands and feedback cross the CAN bus.

Node Controls
STM32G431 Left drive motors
STM32G431 Right drive motors
STM32G474 Front drum (brushless)
STM32G474 Back drum (brushless)
STM32G431 Two linear actuators

The main controller is a deterministic MCU kept separate from the autonomy computer, so the rover stays controllable if the high-level stack stalls.

Autonomy

A ten-package ROS2 workspace: ToF driver, WebSocket bridge, excavation state machine, mission executive, Nav2 configuration, ZED interface, bringup, URDF and elevation mapping.

2027 chassis

The single-mast cantilevered boom is being replaced with a triangulated truss, and each bucket drum gets its own linear rail and motor. Overall length target is 1.4 m, inside the 1.5 m rule envelope.

What went wrong on this project

  1. Symptom

    The excavation section inherited from our reference design had structural problems we could not tune away.

    Root cause

    One cantilevered mast carried both bucket drums on a shared rigid arm and shaft, so every load path ran through the same joint.

    Fix

    Redesigning around a triangulated truss, with each drum on its own linear rail and motor.

    What I do differently now

    When I borrow a design, I review its load paths before its features.

  2. Symptom

    ZED camera topics would not stream in ROS2 while we were debugging on the rover.

    Root cause

    The installed ZED SDK was 5.3.0; the ROS2 wrapper targets SDK 4.x.

    Fix

    Matched the SDK to the major version the wrapper is built against.

    What I do differently now

    Vendor SDK and wrapper versions get pinned together in the setup script, and checked before any field session.

  3. Symptom

    The Jetson Orin Nano fell behind as soon as the ZED depth pipeline and AI inference ran at the same time.

    Root cause

    Each workload fit on its own; together they exceeded what one module could sustain. The ZED 2i is also USB3, not CSI, which added load the original plan did not account for.

    Fix

    Onboard compute moves to a mini PC with an RTX 4060-class GPU. Motion control moved to a separate deterministic MCU, so a stalled autonomy computer cannot take the drive system with it.

    What I do differently now

    I budget compute for the worst concurrent load, not each workload alone, and keep safety-critical control off the machine that does perception.