I build autonomous systems that work effectively alongside humans — predicting how people will move and behave, detecting when they are confused or frustrated, and explaining what a robot is doing and why. My work spans the full stack: from human behavior modeling and real-time multimodal perception to planning algorithms, machine teaching frameworks, and validated user studies with physical robots and vehicles. I am actively seeking research and engineering positions in robotics and autonomous systems.

Research. My work targets efficiently leveraging human inputs for robot learning and operation. This spans making robots human-aware — developing probabilistic pedestrian and driver behavior models that feed into AV motion planners, and real-time multimodal systems that detect user surprise, confusion, and frustration from facial action units and audio features; making robots understandable — designing machine teaching frameworks that explain robot decision-making to individuals and groups using particle filter-based belief models and information-theoretic demonstration selection; and aligning how robots learn from people — formalizing interaction alignment across RLHF and interactive imitation learning paradigms, and evaluating where foundation models still lack the physical intuition (rigid-body reasoning, state understanding, physical plausibility) that humans supply.

Engineering. I design and implement end-to-end systems using ROS across robotics projects: a contact-aware teleoperation system for dexterous manipulation that tracks teleoperator hand pose via MediaPipe/VR, retargets motion via optimal control, and enforces grasp stability through Grasp Wrench Space constraints in a three-stage TAMP formulation, validated in MuJoCo and deployed toward a Kinova arm with a LEAP hand; a safety-aware LLM action planner (Qwen3) grounded in ISO robotics safety standards via a Neo4j knowledge graph with VLM-based scene understanding (Llama 3.2-Vision), validated in PyBullet; a real-time binary XGBoost classifier (79% true-positive rate, ~2.7 s mean detection latency) deployed autonomously in a driving simulator using OpenFace facial action units and OpenSMILE audio features; a Kalman filter-based driver trust estimator with adaptive communication that reduced trust miscalibration by ~40% in user studies; a multimodal engagement prediction system (78% accuracy) for equitable participation in children’s groups using OpenFace and MediaPipe; and fine-tuning of ACT and Diffusion Policy on a low-cost robot arm (LeRobot/SO-101) with an RLHF preference reward model from paired teleoperation demonstrations.

Skills: Python  ·  ROS 2  ·  PyTorch  ·  MuJoCo  ·  Isaac Sim / Isaac Lab  ·  PyBullet  ·  MediaPipe  ·  OpenFace  ·  OpenSMILE  ·  XGBoost  ·  Neo4j / GraphRAG  ·  LLMs / VLMs (Llama, Qwen)  ·  Kalman & Particle Filters  ·  MPC  ·  RLHF & Interactive Imitation Learning  ·  Inverse Reinforcement Learning  ·  ACT / Diffusion Policy  ·  sim-to-real  ·  Hybrid Automata  ·  VR Teleoperation  ·  Franka Panda / Kinova

Education

• Ph.D. in Mechanical Engineering, University of Michigan, Ann Arbor (2021)
• M.S. in Mechanical Engineering, University of Michigan, Ann Arbor (2018)
• B.E. in Production Engineering, Anna University, India (2013)

Research Interests

Human behavior modeling  ·  Explainable AI/robot decision-making  ·  Safety-aware robot planning  ·  Human-robot teaming  ·  Trust in autonomous systems  ·  Multimodal human state estimation