Robotics & ML researcher working on reinforcement learning, sim-to-real transfer, VLA evaluation, and learned control for physical robots.
- extend-SAFE — Reproducible audit of SAFE (NeurIPS 2025) VLA failure detection on OpenVLA/LIBERO. Showed pooled ROC-AUC inflates +0.09 over macro within-task; functional conformal calibration catches only 7–9% of cross-task failures at ≤5% FPR. Public score bundle, CI-replayed reproducibility, 32 tests.
- two_axis_cart_pole — 3D inverted pendulum in MuJoCo with LQR, energy-based swing-up, and PILCO-inspired Gaussian-Process dynamics control.
- PoleBot — Sim-to-real self-balancing robot: MuJoCo model → PPO balance/drive controllers → Jetson Orin Nano deployment.
- MultiCogTrust — Multi-modal trust inference for HRI using cross-transformer attention over speech, text, and gaze (TIER Lab).
- VBEA: A Voting-Based Evolutionary Algorithm for Multi-Objective Path Planning — EMO 2025
Languages: Python, C++
ML & Robotics: PyTorch, Transformers, MuJoCo, Gymnasium, Brax, scikit-learn, SciPy, NumPy
Methods: Reinforcement learning, conformal prediction, CKA, LQR, PILCO, cross-modal attention, model distillation, sim-to-real transfer
Self-study: MIT 6.832 (Underactuated Robotics), UC Berkeley CS 285 (Deep RL), Sutton & Barto