Shreyas Sharma

Shreyas Sharma

I build robots from the firmware up: RTOS tasks on microcontrollers, ROS 2 on the robot, and vision models that have to run on the hardware they're given.

Final-year B.Tech in Automation & Robotics at SRMIST, Chennai. Three internships across UAV firmware, drone sensor fusion and computer vision. Looking for robotics, embedded and edge-AI roles from 2027.

See the projects Email me
Shoulder0.0°
Elbow0.0°
Tool x, y0, 0 mm
TaskStarting
Left alone, the arm takes its blocks apart and rebuilds them as a new shape on the other side. Point to take over, then hold the mouse on a block to pick it up and let go to drop it. On a phone, tap to pick up and tap again to drop.

Work

Internships and research, most recent first.

  1. to Noida, on-site

    Robotics Engineer Intern at Dtown Robotics

    Project: multi-sensor fusion framework for odometry estimation and autonomous drone navigation.

    • Set up perception with an Orbbec Gemini depth camera and an RPLiDAR, and built a map that kept updating and saved itself as new scans and changes in the robot's pose came in.
    • Visualised robot position and orientation in RViz over ROS 2 nodes.
    • Trained on LiDAR and IMU fusion, odometry calibration, and mission planning in QGroundControl.

    ROS 2RVizRPLiDAROrbbec GeminiQGroundControl

  2. to Hybrid

    R&D Software Development Intern at Elcom Innovations

    • Wrote multi-threaded RTOS tasks with semaphores and queues on ARM Cortex-M4 and M7, using FreeRTOS and VxWorks.
    • Developed and flashed firmware for flight-control sensors and actuator drivers, focusing on low-latency sensor acquisition.
    • Supported subsystem integration workflows for a medium-altitude long-endurance (MALE) UAV.

    CFreeRTOSVxWorksARM Cortex-M4/M7

  3. to Remote

    AI Engineer Intern at QID

    • Built a real-time detection pipeline with YOLOv8, OpenCV and ONNX Runtime.
    • Worked on OCR and document understanding, including preprocessing and orientation correction that held up across lighting, rotation and resolution changes.
    • Deployed for low-latency local and cloud (GCP) use, with reproducible Git-based workflows.

    PythonYOLOv8OpenCVONNX RuntimeOCRGCP

  4. to Chennai

    Research project member at SRMIST

    • Worked with a mechatronics team on a 6-DOF robotic arm: object identification and sorting with ROS 2 and OpenCV, plus sensor calibration for manipulation tasks.

    ROS 2OpenCVPython

Projects

The five systems I've put the most into. Each one says what's built and what isn't yet.

Shipped
Works end to end
Active
In development now
Simulation
Verified in simulation only
Prototype
Hardware built, partly working
Concept
Designed, not built

SHADE

Active

Self-Healing Adaptive Deployment Engine

An operations layer for edge AI on a Raspberry Pi 5. It watches temperature, throttle flags, memory, frame rate and camera health, steps the detector down a model ladder before the firmware throttles it, and steps back up only once it's safe. Every decision goes into an audit log you can query.

PythonYOLOv8ONNX RuntimeOpenCVSQLitePrometheussystemd

Source on GitHub

  1. nominal640 px, every frame
  2. frame-skip640 px, every 2nd frame
  3. small-320320 px, every 2nd frame
  4. emergency320 px INT8, every 3rd frame
The degradation ladder from SHADE's default config. Pressure from heat or memory moves it down one rung at a time; calm moves it back up.

Built

  • Decision engine with hysteresis, dwell, cooldown and temperature-slope anticipation
  • Background model preload with a RAM headroom check, swapped in at a frame boundary
  • Watchdog with hang detection, restart on the emergency rung and systemd notify
  • /metrics, /status and a live status page
  • 101 automated tests, plus fault-injection runs on a simulated device

Next

  • Measure on a real Pi 5. No hardware benchmarks are claimed yet, and the thresholds are reasoned defaults.
  • Write up the simulation results as a paper.

Autonomous EV charging pipeline

Active

A 5-DOF arm with an RGB-D camera that finds an EV's charging socket so it can line the connector up with it.

ROS 2 JazzyEVS-YOLORGB-DGazeboRVizURDF

Repository is private

Built

  • Socket detection with a transformer-enhanced YOLO (EVS-YOLO)
  • Arm description ported to ROS 2 Jazzy: URDF, meshes, launch files
  • Arm tested in Gazebo and RViz

Next

  • Connect detection to ROS 2 so the arm aligns with the socket it sees
  • Run it on the physical arm

YOLO visual servoing for ArduPilot

Simulation

A drone that follows a person by turning YOLOv8 detections into yaw, altitude and forward-speed commands.

PythonYOLOv8DroneKitMAVLinkArduPilot SITLGStreamer

Source on GitHub

Built

  • Proportional control from box-centre and box-area error, sent as MAVLink velocity targets
  • Two modes: follow the target, or hold a fixed distance from it
  • Automatic takeoff, altitude safety bounds, and an H.264 stream to QGroundControl on a phone
  • Run against ArduPilot SITL with QGroundControl

Next

  • Fly it on a real airframe

Drone detection and multi-object tracking

Shipped

Detects drones in video and keeps a stable ID on each one, with an alert when a drone stays in view.

YOLOv8ByteTrackBoT-SORTOpenCVCoreMLONNX

Source on GitHub

Built

  • YOLOv8n trained on a custom drone dataset; mAP50 of 0.945 on its validation set
  • ByteTrack and BoT-SORT configs tuned for small, fast targets and crossing paths
  • Evaluation scripts for mAP and for MOTA and IDF1 tracking metrics
  • CoreML and ONNX export for Apple Silicon

Next

  • Train YOLOv8m at 960 px to catch smaller, more distant drones

Behavioural‑cloning robotic arm

Prototype

A servo arm that I'm teaching to repeat motions from teleoperated demonstrations, instead of hand-coding each move.

ESP32PCA9685MG996RTensorFlow

Repository is private

Built

  • Arm hardware: ESP32 with a PCA9685 PWM driver, four MG996R servos and an SG90 gripper
  • TensorFlow experiments on learning motions from demonstrations

Next

  • Collect a proper demonstration dataset and train the policy on it
  • Decide whether inference runs on the ESP32 or a companion board

Smaller builds

Hardware, IoT and vision projects. Open any row for the details.

Open source

Contributor in GirlScript Summer of Code 2026.

Skills

What I've used in the work above. Coursework is listed separately.

Languages

Python, C, C++, Embedded C, Bash. Working knowledge of TypeScript and React.

Embedded

ESP32, ESP8266, Raspberry Pi 5, ARM Cortex-M4/M7, FreeRTOS, VxWorks, PCA9685, UART, SPI, I2C, CAN, MQTT, PlatformIO.

Robotics

ROS 2 (Jazzy), RViz, Gazebo, URDF, ArduPilot SITL, DroneKit and MAVLink, QGroundControl, RPLiDAR, depth cameras.

Vision and ML

YOLOv8, OpenCV, ONNX Runtime, TensorFlow, ByteTrack, BoT-SORT, scikit-learn, NumPy, Pandas. Some PyTorch.

Tools

Linux, Git, Docker, GCP, SolidWorks, Fusion 360, Unity and C#.

Coursework

Control systems, industrial robotics, kinematics, SLAM basics, RTOS and QNX, PLC and ladder logic (lab work on Siemens, Allen-Bradley and Mitsubishi), VFD basics, instrumentation.

Education

B.Tech, Automation and Robotics

SRM Institute of Science and Technology, Chennai. 2023 to 2027.

Certifications

  • Oracle Cloud Infrastructure Generative AI Professional, 2025
  • Oracle Cloud Infrastructure Data Science Professional, 2025
  • Google Cloud: build an LLM and RAG chat application
  • QNX RTOS course
  • Robotics and AI industrial training, My Equation

Let's talk about robots.

I'm looking for robotics, embedded and edge-AI roles, and I'm happy to talk through any of the projects above.

sharma.shreyas.2k5@gmail.com

Based in Chennai, India.