← All work

RoboHax 2026 · Team TheGraduates

Autonomous Driving

Helping an operator choose where a robot goes, instead of every move it makes.

A 90-second match means steering, watching targets and planning at once. We explored a target-selection workflow that gives the person the goal and the robot the movement.

My contribution
Product design, dashboard workflow and the computer-vision approach, alongside Tejas, Eduardo and Roshan.
What we delivered
A physical competition robot and an AI-assisted control exploration. Autonomous performance still needs validation.
Explore the actual build
Our competition robotThe physical build. LoR Core V3 controller, wired motors and four-wheel chassis.

Our competition robot

Four yellow wheels, motors and a wired LoR Core V3 controller on the physical robot

The physical build. LoR Core V3 controller, wired motors and four-wheel chassis.

From steering to selecting

My design question: what should the person control, and what should the system handle?

Manual steering

Watch the field → steer → correct alignment → repeat.

The operator manages strategy and movement together.

Proposed target selection

Select a goal → review tracking → let the robot move → intervene.

Less manual input only helps if state and a stop action stay visible.

Inside the build

Select a photograph to see the design decision behind it.

The arenaOriginal RoboHax photograph: our robot on the Sector Siege field.

The arena

Yellow-wheeled robot with a visual marker among Sector Siege platforms

Original RoboHax photograph: our robot on the Sector Siege field.

What should the operator decide?

Choose the objective, not every motor movement.

We explored visual target selection so the operator could focus on the next goal. Navigation and alignment became separate system responsibilities.

Design tradeoff

Less steering requires clearer feedback about what the robot sees and intends to do.

Look closer at the hardware

Select a hardware detail to bring its photograph into view.

Inside the proposed control loop

RealSense RGB-D camera → Python / OpenCV / YOLO perception → operator-selected target → navigation → LoR Core V3 / ESP32 → motors. IMU feedback adds orientation information.

Original photos document the build and event. GPU compute and VLA are exploration areas; voice is proposed. No autonomous competition score or trained-policy benchmark is claimed.

What connects the parts

Our approach separated perception, orientation and motor control so failures could be traced to a layer. Camera and GPU details below describe the system exploration, not hardware visible in every photograph.

Vision + sensing approach

  • Python
  • OpenCV
  • YOLO
  • Intel RealSense RGB-D
  • IMU

Robot control

  • LoR Core V3 / ESP32
  • Arduino
  • C++

Compute + model exploration

  • GPU compute
  • VLA exploration

Built, with more to prove

The robot and arena are real. The next proof is whether target selection leads to reliable movement, not just a convincing detection on screen.

The first next test

Measure target-reaching success and how often an operator intervenes. Validate that loop before adding VLA learning or voice commands.

Next projectMedBot