Robotics M.S. student at the University of Minnesota. Background in physics and mechanical engineering from BITS Pilani '25.
Currently in the RPM Lab working on in-context imitation learning for bimanual manipulation on dual UR5e arms, and building multimodal data collection pipelines with ROS2.
Previously at the Indian Institute of Science (Stoch Lab), where I trained and deployed RL locomotion policies on a 70 kg custom quadruped.
- GRoQ-LoCO: Generalist Quadruped Locomotion via Offline Datasets — arXiv:2505.10973
| Project | What it does |
|---|---|
| Barrier-Loco | Asymmetric actor-critic with barrier critic for quadruped locomotion on rough terrain |
| multimodal-robot-data-collection-pipeline | ROS2 pipeline recording RGB-D, GelSight, and gripper state on dual UR5e arms |
| ip (bimanual branch) | Extending Instant Policy to bimanual manipulation with per-arm graph representations |
| vlm-grasp | Language-guided grasp detection with fine-tuned Qwen3-VL-8B |
| dream-team-research-automation | Multi-agent research system with dynamic specialist recruitment and safe code execution |
| cv5561-f25-team-asa | Class-aware seam carving combining DINOv3, Depth Anything 3, and RGB-VST saliency |
| Project | What it does |
|---|---|
| worldvalues-qwen-finetuning | Data synthesis for LLM fine-tuning on WorldValuesBench with Qwen3 |
| isaac-gym-server-gui | Run Isaac Gym with live GUI on headless HPC clusters via VNC |
| arithmetic-transformer-experiments | Small transformers for elementary arithmetic with abacus embeddings |
| planning-optimal-grasps-html-demo | Interactive explainer for Ferrari and Canny's 1992 optimal grasp planning |
| dangerous-dave-mod | Dangerous Dave on Windows 11 with gameplay cheats via bootstrapped DOSBox |