Utkarsh Gupta

Robotics Engineer focused on Perception & Autonomy

I work across both hands-on robotic systems and simulated environments, with a focus on perception, autonomy, and end-to-end decision-making models.

My experience spans sensor fusion, 3D reconstruction, navigation, and vision-language-action (VLA) models. I have deployed perception and planning stacks on real robots and built scalable simulation pipelines for real-to-sim transfer and policy training.

I enjoy working across the full robotics stack—from data and models to deployment and evaluation— particularly in research-driven and applied autonomy settings.

Utkarsh Gupta

Experience

Research Assistant

PSI Lab, USC

Oct 2025 – Present

  • Built simulated data pipelines in IsaacSim and MuJoCo for real-to-sim VLA training.
  • Implemented domain randomization and force-feedback for diverse data generation.
  • Integrated heuristic and VLA policies for trajectory collection.
  • Open-source real-to-sim data collection framework: OpenReal2Sim

Research Assistant

CPS Vida Lab, USC

Nov 2024 – Nov 2025

  • Worked on interpretability and safety analysis for VLA models.
  • Improved action success rate by 20% using human-in-the-loop policies.
  • Developed imitation learning strategies for multi-agent path finding.

Robotics Engineering Intern

ERIC Robotics

May 2025 – Aug 2025

  • Designed real-time railtrack and 3D safety algorithms for AGVs.
  • Reduced AMR localization time by 40% using optimized LiDAR pose estimation.
  • Deployed CUDA-accelerated perception pipelines on Jetson AGX Orin.

Education

M.S. in Computer Science

University of Southern California

Aug 2024 – May 2026

GPA: 3.9 / 4.0

Focus on robotics, perception, and autonomous systems.
Coursework: DL for Robotic Manipulation, Robotic Perception, Autonomous Systems, Robotic Learning

B.Tech in Computer Science

MIT Pune

Dec 2020 – Aug 2024

GPA: 3.8 / 4.0

Worked on applied ML and optimization. Participated in AI lab research and geospatial analysis projects.

Projects

Digital Twin via 3D Reconstruction

Photogrammetry-based 3D reconstruction pipeline for high-fidelity digital twins, including partial-view scans using Poisson surface reconstruction.

RL-Based Stealth Navigation

Multi-agent reinforcement learning environment with partial observability, achieving 85% task success under adversarial constraints.