Autonomous Urban Mobility for the Blind
Undergraduate Researcher
I handle the mechanical overhaul and accessibility redesign for the platform, rebuilding the steering motor mount, brake servo mounts, and drivetrain so the autonomy stack has reliable hardware to drive. We took over the hardware from a senior capstone, and the prototype barely functioned mechanically; steering would lock mid turn and the chain slipped constantly, so any autonomy on top was building on commands that couldn't execute. I also redesigned the chassis for blind accessibility, including capping sharp edges at touch height and adding entry handles so a user can locate and enter the bike without sight.
The Problem
The lab works on assistive technology for visually impaired users, and this platform was inherited from a senior design capstone that prototyped a semi autonomous recumbent tricycle. The hardware base is a Jetson Orin Nano reading 3 OAK-D depth cameras and VectorNav IMU over ROS, driving a 12V DC steering motor and servo actuated brakes. My work is the mechanical overhaul and accessibility redesign to bring the platform to a state where it can actually be ridden, and where the autonomy stack has reliable hardware to drive. The computer vision stack, including camera calibration and perception, is handled by others in the lab.
Mechanical Overhaul
The capstone left the platform in a state where the autonomy stack had nothing reliable to actuate. Steering would lock mid turn, the pedaling chain slipped under load, and the brake servo couldn't consistently pull the brake handle to full engagement. Before anything autonomous could be layered on top, the mechanical system needed to actually execute the commands it was given.
The steering motor was mounted off axis from the tie rods it pushes, so its linear travel bound against the linkage instead of translating cleanly into a steering angle at the wheels. I milled a new aluminum motor mount that aligns the motor with the tie rod axis, which eliminated the binding and made steering responsive to motor commands. The brake servo mount was rebuilt so the cable pull reaches full engagement on the brake handle, and the pedaling chain was re-tensioned with a proper guide after re-aligning the sprockets.
Accessibility Redesign
The tricycle also had to work for a user who can't see it. I smoothed or capped sharp edges along the aluminum cowling at head and hand height, since a blind rider locating the bike by touch would otherwise catch themselves. I added a guide handle near the seat so entry doesn't require feeling across the chassis for a reference point, and repositioned exposed fasteners that could snag clothing during pedaling.
User Trials
The hero video at the top of the page is from internal team testing; the bike ran end to end with no one on the frame and no manual override in use, confirming the platform works on its own. With that baseline established, we ran trials with blind users to collect feedback on ergonomics, pacing, and feedback behaviors.
ML Integration
The routing framework from my other work at the lab may end up deployed on this platform down the line. The local versus cloud tradeoff applies to the bike directly; running the cloud model on every frame is too slow for safe obstacle response in traffic, and running only the local model risks missing the kinds of edge cases a blind rider cannot recover from visually. Integration isn't committed yet, since the mechanical base needs to be stable first.