Projects

Research

The Price of Paranoia: Robust Risk-Sensitive Cooperation in Non-Stationary Multi-Agent Reinforcement Learning: AAMAS 2026 (ALA Workshop). Developed RATTL, an algorithm that uses a closed-form trust factor to resolve the EVaR Paradox in MARL. By targeting policy gradient update variance rather than return distributions, we provably expand the cooperation basin in coordination games.

Reinforcement Learning & Robotics (StochLab, IISc): Extended the VipLoco visual world-model framework for height-conditioned locomotion and low-clearance traversal. Designed composite reward formulations for gait stability and validated zero-shot policy transfer between Isaac Lab and MuJoCo for deployment on quadruped hardware.
Intelligent Radio Sensing with Multi-Armed Bandits: 10th Place at Inter IIT TechMeet 14.0 (Arista Networks Challenge). Designed a UCB-based Multi-Armed Bandit framework with EWMA rewards and built a Model-Based RL system using a GNN environment model to evaluate action safety.
Primal-Dual Optimization of Systems with Hard Feasibility Constraints: Formulated a Lagrangian-based OFDM scheduler enforcing strict QoS guarantees under hard minimum-rate constraints with a shadow price capping mechanism to handle infeasibility.
Single Image Super Resolution: Implemented and trained a Diffusion Probabilistic Model (DDPM/SR3) in PyTorch, using Resize-Convolutions and Self-Attention to eliminate checkerboard artifacts and improve texture coherence. GitHub
Fault-Tolerant Quadrotor Control: Bronze Medal at Inter IIT TechMeet 13.0 (IdeaForge Challenge). Developed an autonomous motor failure detection system with a fault-tolerant PID controller in C++/PX4, achieving 97.5% detection accuracy and 100% stable landings. GitHub