Bavin Saravanan

I’m a Master’s student in Mechanical Engineering at Carnegie Mellon University and a research assistant at the CMU Robotics Institute, working in AirLab with Dr. Sebastian Scherer and in the Legged Systems Group with Dr. Hartmut Geyer. I focus on building intelligent autonomous systems that operate reliably in real-world environments. My work lies at the intersection of control theory and learning-based approaches, particularly reinforcement learning and multi-agent coordination, with an emphasis on safety, robustness, and decision-making under uncertainty. I’m especially interested in bridging the gap between theory and deployment, developing methods that not only perform well in simulation but can be transferred effectively to physical systems. This includes tackling challenges in autonomous navigation, system reliability, and scalable coordination among multiple agents.

I hold a Bachelor's degree in Mechanical Engineering from SSN College of Engineering (Chennai, India). My foundation in robotics and control has been built through hands-on research and independent projects, shaping my approach to designing systems that are both principled and practical.

Email  /  CV  /  Github

Publications

Developed AM-Bench, a modular benchmark for aerial manipulation policy learning across contact, transport, and constrained-interaction tasks, evaluating how embodiment, low-level control, disturbances, and policy design interact under multirotor dynamics.

Compared EDM, powder-mixed EDM, and laser beam machining on Cu-Al-Fe shape memory alloy using SEM, AFM, and roughness profiling, identifying PMEDM as the most effective process for reducing surface roughness.

Projects

Introduction

Robots that combine walking and manipulation could transform hazardous work, disaster response, warehousing, eldercare. Existing approaches require expensive data, lack interpretability, or fail to generalize.

Design

A three-layer modular architecture:

• Perception - RGB-D, LiDAR, and IMU for 3D scene understanding. A VLM enables open-vocabulary object grounding; a ReID module tracks objects across occlusions.
• Neural - VLM grounds language instructions in the scene, locating any named object without retraining.
• Control - PPO policy on a Unitree G1 (29 DOF), trained in Isaac Lab with a two-phase reward transitioning from navigation to manipulation at 0.75 m.

Evaluation

Policy transferred successfully to the physical robot with consistent phase transitions. Limitation: grasping success rates dropped due to unmodeled contact dynamics.

Conclusion

Modular loco-manipulation is achievable without large-scale data. Next steps: domain randomization and tactile feedback to close the sim-to-real gap.

Introduction

Search engines return links; LLMs hallucinate. This project combines both — grounded, accurate answers by anchoring LLM output to retrieved content.

Design

• Retrieval - Queries embedded via nomic-embed-text, matched against a custom SQLite vector database built from crawled source URLs.
• Generation - Top-k chunks passed as context to a Llama-instruct model, grounding answers in retrieved documents to reduce hallucination.
• Architecture - React frontend decoupled from Flask backend via REST API; embedding model, vector store, and LLM each independently swappable.

Evaluation

Outperformed keyword search on factual precision; full component swaps validated in under ten minutes. Limitation: retrieval recall drops when source coverage is sparse.

Conclusion

RAG effectively grounds LLM output. Reliable AI systems demand disciplined architecture, not just model selection.

Introduction

Most RL practitioners use libraries as black boxes, limiting their ability to debug or adapt. This project implements 9 core RL algorithms from scratch — directly from the original papers.

Design

• Value-based - Q-Learning, DQN, Dueling Double DQN
• Policy-based - REINFORCE
• Actor-critic - A2C, DDPG, PPO, SAC, TD3

Each is a single self-contained file, heavily commented, clarity over abstraction.

Evaluation

Trained and evaluated policies on benchmarks including Ant, HalfCheetah, Hopper, and BipedalWalker, for both discrete and continuous control.

Conclusion

Implementing from paper to code reveals gaps that using libraries never exposes.

Introduction

Tamil is spoken by 80 million people yet severely underrepresented in AI research. This project builds a Tamil language model from scratch — proving capable LLMs are achievable for low-resource languages.

Design

• Pre-training - Trained on mC4, news articles, storybooks, and Wikipedia using data parallelism across dual A6000 GPUs (96 GB VRAM).
• Fine-tuning - SFT on Tamil conversation datasets for coherent multi-turn dialogue.

Evaluation

Model produced coherent Tamil dialogue on held-out prompts. Limitation: scale — a larger corpus would improve factual coverage.

Conclusion

Capable low-resource LLMs are buildable with constrained compute and careful data curation.

Experience

Research

• Informative Path Planning - Contributed to AirLab's multi-agent UAV exploration system built on the AirStack autonomy stack. Integrated RayFronts for open-vocabulary semantic mapping, enabling class-based semantic exploration where UAVs prioritize regions by object category, and designed a general swarm interface for scalable, coordinated autonomous exploration.
• Multi-Agent Pursuit-Evasion - Formalizing an intruder-evader swarm framework and its underlying theory. A privileged expert uses Monte Carlo Tree Search (MCTS) with full state information and a learned game dynamics and value function to generate plans, while an onboard flow-matching policy plans from local measurements for decentralized execution.

• RL Prosthesis Policy - Developed and deployed a reinforcement learning control policy for the CMU leg, enabling online adaptation and stable, natural walking. Trained the policy with diverse motion priors to capture a range of human gait patterns; ongoing work integrates terrain maps for more generalized prosthesis locomotion.
• Prosthesis Hardware Redesign - Redesigned the University of Michigan Open-Source Leg (OSL), replacing its actuators with lightweight CubeMars motors and redesigning the entire motor mount for the new drivetrain.

Professional

• Decentralized VTOL UAV Swarm (Cactus Swarm) - Developed a decentralized multi-agent VTOL swarm demonstrating flocking, foraging, and cooperative payload delivery. Designed a novel vector-field-based guidance algorithm for smooth, coordinated swarm navigation, and built a lightweight inter-UAV communication stack and a Jetson Xavier NX-PX4 control pipeline via ROS2 and Micro XRCE-DDS. Achieved fully onboard autonomous flight up to 2 km altitude and 12 km range.
• GPS-Denied Autonomous Quadrotor (Bumblebee) - Developed a palm-sized quadrotor (Jetson Orin NX) for autonomous navigation in forests, combining Isaac ROS SLAM for real-time 3D mapping, A* planning over local obstacle maps, and a Lissajous search strategy. Implemented real-time detection, tracking, and Siamese-network-based target verification in cluttered environments.

Teaching Experience

Support students in applying ML and AI methods to engineering problems through office hours and assignment feedback.

Mentored project teams on experiment design, model implementation, evaluation, and technical communication.

Helped students connect robustness, uncertainty, fairness, and evaluation concepts to practical ML workflows.

This website is based on template at here.