Work No. 042022

Mitra Robot

Formerly Invento Robotics

Overview

Mitra Robot, formerly Invento Robotics, builds commercial humanoid robots with large language models built in, for customer engagement, healthcare assistance and visitor management.

Role
Software Engineer Intern
Employment type
Internship
Dates
Aug – Dec 2022
Team
Two interns, reporting to the Chief Software Architect
Company
Mitra Robot (formerly Invento)

Delivered

Live · senior care, Texas

Fall detection

< 20 sto alert medical staff

  • Real-time fall detection model: fine-tuned vision transformers with OpenAI CLIP, in TensorFlow and PyTorch
  • Live on a ROS2 robot in a senior care facility in Texas, USA
  • Recognises falls in low light, too
CLIP · ViTs · OpenCV · MediaPipe

Gesture recognition

Older residents' voices can be soft or broken, so speech alone fails. Simple gestures say what they need.

36 gestures in all, for example:“I want food”“Show me my family's latest photos”“Play some music”

  • Ran a data collection event with young and older people
  • Collected, cleaned and labelled every training video
  • Fine-tuned CLIP-based models and other vision transformers, with OpenCV and MediaPipe, to recognise each gesture

20,000+gesture videos

36gestures

25+participants, young and old

The dataset on Kaggle
Data → train → evaluate → deploy

The ML pipeline

  • Worked across the whole AI/ML pipeline, from data to deployment
  • Ran model evaluations, testing each model against how the robot actually performed
  • Wrote storage optimisation algorithms to store large amounts of robot-collected video data
WebRTC · AWS · GCP

Robot operations

  • Fixed WebRTC bugs
  • Handled AWS and GCP
  • Kept the robots running smoothly
Internship certificate
Python / PyTorch / TensorFlow / ROS2 / Computer vision / OpenCV / MediaPipe / CLIP / Vision transformers / Statistical modelling / Model evaluation / Dataset collection / WebRTC / AWS / Google Cloud / Human–robot interaction / 

Stack: Python, PyTorch, TensorFlow, ROS2, Computer vision, OpenCV, MediaPipe, CLIP, Vision transformers, Statistical modelling, Model evaluation, Dataset collection, WebRTC, AWS, Google Cloud, Human–robot interaction

Next · ProjectsGameNight →
Lekhana Kalyan Raj
RésuméLinkedInGitHub