Noushad Sojib

Robot Learning Engineer building robust policies from imperfect human demonstrations. Focused on imitation learning, VLA models, and diffusion policies for real-world robotic systems.

M.Sc. in Computer Science from the University of New Hampshire (Cognitive Assistive Robotics Lab). B.Sc. from SUST, where I founded a robotics club and built humanoid robots from scratch.

Currently seeking roles in robot learning and real-world robotic systems.

Email  /  GitHub  /  Google Scholar  /  LinkedIn

profile photo

Robot Learning

Projects focused on learning robust robot behaviors from human demonstrations, spanning imitation learning, VLA models, and diffusion policies.

robot learning robot learning robot learning

Real-World Robot Learning with Diffusion Policies and VLA Models



Applying state-of-the-art robot learning frameworks—including Diffusion Policy, π0.5, OpenVLA, and Groot N1.5—to train manipulation policies from human demonstrations. Focus on generalizing across tasks and environments with minimal data, targeting real-world deployment on physical robot platforms.



Robot From Scratch

I started building robots from scratch out of curiosity—there were no humanoid platforms available to learn from where I grew up. This led me to found RoboSUST, where I led a team to design, build, and deploy multiple robotic systems using low-cost hardware and self-developed control pipelines.

project image

Ribo — 24 DOF Humanoid Robot



Designed and built a full humanoid robot capable of upper-body manipulation and human-interactive behaviors.

Role: Team Lead — hardware, software, and interaction interface

Key Contributions: Led hardware and software development of a 24 DOF humanoid platform. Implemented control for coordinated arm and hand motion. Designed user-facing interaction interface.
project image

Lee: A biped walking robot



Built a biped robot focused on achieving stable walking with minimal hardware cost.

Role: Team Lead — mechanical design, gait control, and software

Key Contributions: Designed mechanical structure for balance and locomotion. Implemented basic gait generation and control. Optimized for low-cost components.
project image

Kiddo



Interactive educational robot designed to engage children through programmable behaviors—built and validated in both simulation and physical hardware.

Role: Solo Designer & Developer




Hardware Design

Building robots from scratch taught me that good software needs good hardware. Along the way, I designed several embedded devices—a few of which landed in peer-reviewed venues and are now in active use on research robots.

project image

3Wheel Mouse



Three-wheeled input device that enables efficient, versatile non-visual computer interaction for blind users.

Role: Designer & Prototype Builder — published at ACM UIST 2024

Islam, Md Touhidul, et al. “Wheeler: A three-wheeled input device for usable, efficient, and versatile non-visual interaction.” ACM UIST 2024. Paper & Video
project image

Charging Dock



Robust, low-cost autonomous charging dock for mobile robots—enabling continuous operation without human intervention.

Role: Designer & Prototype Builder — demonstrated live at IROS 2023, deployed on Hello Stretch

Live demonstrated at IROS 2023. An extended version is actively used with the Hello Stretch robot. See example
project image

Lowcost Braille Display



Low-cost single-cell Braille display that makes digital Bangla text accessible to visually impaired readers.

Role: Designer & Prototype Builder — published at IEEE ICBSLP 2018

Sojib, Noushad, and M. Zafar Iqbal. “Single cell bangla braille book reader for visually impaired people.” IEEE ICBSLP 2018. Paper



Research

How can robots learn reliably from imperfect human demonstrations? I develop methods that enable robots to extract safe, generalizable behaviors from noisy or erroneous data, bringing imitation learning closer to real-world deployment.

project image

Self Supervised Detection of Incorrect Human Demonstrations: A Path Toward Safe Imitation Learning by Robots in the Wild


Noushad Sojib, Momotaz Begum
IROS 2024, 2024

Problem: Human demonstrations are often noisy and degrade policy learning. Approach: Proposed BED, a self-supervised method to detect and filter incorrect demonstrations. Result: Enables robust policy learning from real-world, imperfect data. Validation: RoboSuite simulation + real Sawyer robot arm deployment.
project image

Self-Supervised Visual Motor Skills via Neural Radiance Fields


Paul Gesel, Noushad Sojib, Momotaz Begum
IROS 2023, 2023

Problem: Learning visuomotor policies without labeled data. Approach: Combined NeRF-based scene representation with keypoint correspondence for self-supervised learning. Result: Learns manipulation skills directly from raw visual input with improved generalization.

Design and source code from Jon Barron's website