Ruthwik Dasyam

Hi, I'm Ruthwik...

I am graduating from UMD College Park this summer. I spent my two years learning a lot here—won competitions, broke robot parts, wrote thousands of lines of code, pitched ideas, skipped meals, pulled all-nighters, and took on extra credits - Now, am graduating with invaluable skills and experiences. From brewing the best coffee to programming an entire software stack, I've improved at everything. And, I will continue teaching robots to do great things. Ultimately, it's about robots and humans living every day with happiness and excitement.
I am looking to meet and join roboticists who are on the same path—making every day interesting.

Twitter | LinkedIn | Github | GoogleScholar | Resume

NEWS

Mar' 25 - Patent Grant for my undergrad Thesis 🔗
Mar' 25 - Submitted a paper to IROS 2025 🔗
Jan '25 - Recipient of 'Pathways for PhD' program grant - Maryland Robotics Center
Oct '24 - Attended NVIDIA AI Summit - Washington DC
May '24 - Best project among 100+ robotics Grad students - Path planning course
May '24 - Top 2 in TurtleBot Path Planning Competition
Dec '23 - Chosen as best project - Robot Modelling Course
Sep '23 - Started at the University of Maryland, College Park for Masters in Robotics
May '23 - Graduated from IIITDM Chennai, Silver Medallist
Feb' 22 - Published a paper at IEEE OCEANS Conference on Autonomous Underwater Vehicles 🔗

SKILLS

Great at Python, Numpy, Pytorch

Robotics : ROS1&2,
Gazebo, MuJoCo, Issac Sim
VLM, OpenCV, YOLO, SLAM
NVIDIA CUDA

Reinforcement and Imitation Learning
LLM-based task sequencing
Foundation Models

RESEARCH

Graduate Researcher, Prof. Pratap Tokekar @ UMD, RAAS - Robotics Algorithms & Autonomous Systems
Aug 2024 - Present

Language-Conditioned Mobile Manipulator

Built a Software package from scratch for the custom-made manipulator (Scaled UR3e), with an RL wrapper to train and test reinforcement learning algorithms.
Teleoperation: Designed hardware setup and software code to teleoperate the active arm with passive and collect data.
IL + RL: Testing various IL-based RL baselines, such as IBRL, to perform pick and place tasks.
LLM's integration to analyze user queries and determine the goal position for the turtle bot.
Generalist Policies to enable adaptive, zero-shot task execution in manipulation tasks.
Fail-safe methods: Exploring methods to ‘try again’ or ‘ask for help’ that ensure successful task execution, adapting to physical and environmental constraints.

RL+IL Methods, PyTorch, Inverse Kinematics & Dynamics, SLAM, MuJoCo, ROS1&2, Generalist models pretraining and testing, LLM task sequencing, 3D printing, Dynamixels, NVIDIA Jetson Nano, Python, Realsense Cameras, OpenCV

Click Here for Media

Preference-Based RLHF with Vision Models for Manipulators 🔗

RLHF + VLM: Integrating GPT4V and Gemini into RL for ranking observation images.
RLHF+VLM+Sketch: Combining trajectory sketch images with observations for VLM to choose for better reward function generation.
RL Training on Clusters: Training and evaluating models on large-scale GPU clusters such as Nexus and Lonestar6, leveraging various NVIDIA GPUs to optimize RL training efficiency.

Using VLM Models- GPT-4V&CLIP, Pytorch, OpenCV, Algorithms from StableBaselines3, Metaworld simulations, GPU Clusters

VARP: Reinforcement Learning from VLM Feedback with Agent Regularized Preferences

A Singh, A Bhaskar, P Yu, S Chakraborty, R Dasyam, A Bedi, P Tokekar
Submitted to IROS 2025

Graduate Researcher, Prof. Nikhil Chopra, Maryland Robotics Center @ UMD
May 2024 - Aug 2024

Multi Drone Cooperative Transport 🔗

Framework: Developed a software package to train and test Drones with RL algorithms in MuJoCo.
Leader-Follower System: Follower drones are trained on DDPG and SAC using OpenAI Gymnasium to support a leader drone operating on PID control.
Sim2Real: Trained models are further evaluated with multiple Crazyfly2 drones, achieving a success rate of 95% with two drones, 75% with three drones, and 88% with four drone configuration systems.

OpenAI Gymnasium, MuJoCo, Crazyfly2 drones, PyTorch, Python

CODING

AI Assistant - Multimodal RAG System 🔗

Query relevant data retrieval from text and image files using FAISS Index and Colpali with QVec VLM model from Hugging Face for indexing.
Multimodal User Interaction: Created speech and text-based interaction using Google STT and TTS APIs to convert voice input from the user to text and output text from Mistral LLM to voice.
Improved Speed: reduced query response time to under 5 seconds.

PyTorch, Hugging Face Models

Teaching a Humanoid Robot to walk 🔗

Stable Baselines3: Designed models using PPO, SAC, and A2C to train a custom humanoid robot model to walk.
OpenAI Gym + MuJoCo: SAC performed best - 1e8 timesteps. BRAX+MJX: PPO performed best - 3e7 timesteps.
Hyperparameter tuning: Formulated and optimized reward models achieving natural gait for an 18 DOF system.

PyTorch, Python, OpenAI Gymnasium, StableBaselines3, MuJoCo, Solidworks

Novel Path Planning Algorithm - RURRT* 🔗

10x Faster: Finds optimal paths 10x faster than the traditional RRT* algorithm.
Reduce time and computation: Algorithm builds upon RRT* by utilizing paths from previous trees, converging quickly regardless of query changes in the same map. Reduced path compute time from 8.2s to 0.7s using RU-RRT.
Exploring ways to implement in 3D and adapt to dynamic obstacles.

Python, Pygame

TurtleBot Race

🏆 2nd place among 100+ Robotics Masters Students - finishing the task in 33 Seconds
Performed path planning using A star algorithm along with proportional controller to navigate using Turtlebot3 Waffle.

Programming of TurtleBot Waffle, RRT*, ROS2, PathPlanning Methods, PID Control System, LIDAR plugin, Gazebo

Autonomous Driving 🔗

Follows the papers, while reacting to STOP signs, avoiding dynamic obstacles. Thus, autonomously navigates throughout the path.

Computer Vision, Hough Transformation, Harris corner detection, Stitching Images, RANSAC for feature matching, Homography Techniques, Contour Mapping, Dynamic obstacle detection, YOLOv5

Transformable Robotic System 🔗

A robot that can transform between a wheeled vehicle - Underwater vehicle - drone.
Executed inter-environmental pick-and-place tasks using the vehicle’s integrated 4-DOF manipulator and vacuum gripper.
Simulated using ROS2 in Gazebo and performed inverse kinematics with manipulator.
🏆 Rewarded extra points, among 100+ graduate students.

Forward and Inverse Kinematics, ROS2, Robot Modelling, URDF, Gazebo, Solidworks

Hybrid Manipulation

Modeling and control of a hybrid rigid-soft robot - rigid and soft links serially attached.
Built closed loop control system for this manipulation system, which sync between soft and hard links to pose for desired end effector position.

Motion capture system, Adaptive control, Arduino Programming, MATLAB Simulations, PID tuning, Closed loop control with MoCap, Soft robot prototyping and control.

Path Planning and Navigation using LIDAR 🔗

Autonomous and teleop driving in a Gazebo simulation with a custom modeled 4-wheeled vehicle.
The robot senses the environment around it, using LIDAR, and computes the linear and angular velocities required for it to travel without collision. This robot is placed in a custom made Infinite shaped path, that can make the robot traveling in a loop, in its path without collision

LIDAR, Rviz, ROS, Gazebo, PID Control

Autonomous Underwater Vehicle Society

Design engineer and Team Lead for a society established to build Autonomous Underwater Vehicles for International Underwater Competitions.

Architected navigation software for perception-based navigation controlling six thruster vectored configuration
Tested vehicle’s performance in an underwater environment and improved its performance from 35% to 98% by enhancing perception algorithms.
International Competitions 🏆: Achieved top performance in SAUVC 2022 (Singapore) and MATE ROV 2022, California (Online)

ROS, Gazebo, OpenCV, Python, C++, Underwater Perception, 3D printing, Waterproofing, CAD, CAE

Design and Performance Analysis of Bio-Inspired Remotely Operated Robot (With emphasis on MATE ROV explorer challenge)

JR Sukesh, R Dasyam, S Muthu, V Iyengar
IEEE OCEANS 2022

Page updated

Google Sites

Report abuse