Abirath Raju

Social navigation of Humanoid robot Digit (Perception and RL)

As a part of my capstone project, I developed an end-to-end navigation pipeline that directly maps latent features to bipedal robot actions. I used collision regions, emotion-related discomfort zones, and human-human interactions to derive the latent features. I developed an actor-critic reinforcement learning framework to train with full-body bipedal robot dynamics, effectively mitigating model discrepancies introduced by reduced order models (ROMs).

System integration of Humanoid robot Digit

As a part of the LIDAR Lab, I worked on system integration of a humanoid robot (Digit). I worked to help Digit navigate through an environment with humans and avoid collisions. My work involved using the Vicon Mocap system to detect humans in the vicinity and using the Zonotope-based model-predictive controller to plan Digit's trajectory.

Work published at T-ASE (IEEE Transactions in Automation Science and Engineering ).

Foundational models to understand failure recovery in robots

As a part of the Deep Learning for robotics coursework with Dr. Danfei Xu, my team and I explored VLM's to create open-world scene graph representations and used an LLM to identify robot failure given the event summary of the task performed by the robot. We evaluated our methodology on both AI2Thor simulator data and real life teleoperation data collected on the Stretch robot.

Visit GitHub →

Turtlebot3 Burger

As a part of the Robotics research course (7785) at Georgia Tech, I developed algorithms using Python3, ROS2 and Gazebo to help the Turtlebot3 robot perform tasks like following an object, avoiding obstacles and navigating to waypoints, path planning with Navstack and maneuvering an obstacle course.

Visit GitHub →

Biomechanics and machine learning

I was selected to be a MITACS Globalink Research Scholar and had the privilege to conduct research at the Human Performance Lab, New Brunswick, Canada. I worked under the supervision of Dr.Victoria Chester where I researched and developed various ML and DL models including CAT Boost, SVM with RBF kernels, KNN algorithms, LSTM’s and CNN’s for the binary classification problem of separating Autism vs Control class groups using multi-segment foot data.

Visit GitHub →

Drone simulation and path planning

Worked under the supervision of Dr.Deepak Mishra at the Indian Institute of Space Science and Technology. I modeled a Quadcopter UAV in Gazebo integrated with Hokuyo LiDAR and Ardupilot controller. Created a custom environment with obstacles and used the RRT* algorithm to navigate the drone from start to goal.

Visit GitHub →

Myoelectric pattern recognition

Worked under the guidance of Prof.P.A.Karthick on researching and developing CNN neural networks for classifying hand sEMG signals to be used in prosthesis for amputees,in collaboration with IIT Madras. I used the NINAPRO database, performed signal processing techniques using MATLAB and fed into neural nets. Achieved accuracy of 70 percent on a 10 class classification problem to classify hand movements like thumb flexion,extension etc.

Visit GitHub →

KUKA robot vision

Worked under the guidance of Prof. Dr. S.K Pal. I built a colour detection software using OpenCV, which is implemented in a 6 D.O.F robotic arm at the Centre of Excellence,IIT Kharagpur. I also worked with TCUP Edge (IOT platform) to interface with the sensors of the FSW robot and building robust machine learning algorithms to control force and torque parameters.

Visit GitHub →