Hi, I'm Kartik Virmani

Deployed PyTorch reinforcement learning policies on real quadrotors using ROS 2, Betaflight MSP, and ESP32-based CRSF/RC override with explicit safety and authority handoff mechanisms.

Built a complete LiDAR-only autonomy stack using SICK PicoScan, including SLAM, localization, obstacle avoidance, and ROS 2 navigation for robust real-world deployment.

Developed bimanual manipulation environments in IsaacLab and explored flow-matching policies for coordinated catching, throwing, and interaction without explicit trajectory optimization.

Built a complete ROS 2 autonomy pipeline integrating localization, navigation, and motion planning (RRT, Pure Pursuit, MPC) for holonomic mobile robots.

Implemented 3D path planning, vision-based localization, and PID/MPC controllers in simulation and hardware-in-the-loop testing environments.

Built a decoder-only Transformer with causal masking, BPE tokenization, and autoregressive sampling for LLM. Implemented a VAE and a DDPM for MNIST, logging ELBO terms, loss curves, and denoising trajectories.

Integrated GroundingDINO, SAM, and Kalman Filters into ORB-SLAM3 for dynamic object masking, significantly improving trajectory accuracy and map density.

Implemented Stein Variational Guided MPPI on the F1/10 platform, achieving robust high-speed trajectory tracking at 50 Hz update rates.

Fused IMU, GPS, LiDAR, and encoder data using an Unscented Kalman Filter to minimize localization drift over extended trajectories.

Implemented NeRF in PyTorch for novel-view synthesis with positional encoding, hierarchical sampling, and comprehensive PSNR/SSIM evaluation metrics.