Ian Pedroza

Hello! I'm a fourth-year undergraduate in the Jerome Fisher M&T Program at Penn, and a Robotics Accelerated Master's in the Penn GRASP Lab. My work and interests are towards generalist robot learning policies for fine-grained manipulation tasks.

At Penn, I'm a research assistant in the Perception, Action, and Learning (PAL) Group , under Professor Dinesh Jayaraman, where I work on imitation learning and foundation models for robotic manipulation. I previously interned at Daxo Industries developing imitation learning pipelines for agricultural robotics.

CV  /  Email  /  LinkedIn  /  Github

Research & Projects

I strongly believe in the need for generalizable, flexible policies that can enable adaptation to various environments, embodiments, and task specifications. My work focuses on imitation learning, reinforcement learning, and foundation models for robotics, with implementations across various real robots and simulation environments.

Imitation learning from egocentric human web videos for zero-shot deployment on household robotic manipulation tasks.
• Conducted experimental trials with a Franka Emika robot arm on everyday tasks like opening/closing drawers, pouring, stirring, etc.
• Wrote the software for deploying policies on a Widow-X-250 robot, then showcased cross-embodiment generalization through experimental trials

Multi-level controller, featuring a high-level footstep trajectory optimizer and a low-level reinforcement learned joint controller to traverse extreme terrain.
• Developed a terrain-aware footstep planner in PyDrake to solve direct collocation problems for valid footsteps
• Trained a low-level reinforcement learning policy in Isaac Gym to enable robust execution of planned footsteps

Vision transformer-based dynamics model to predict deformable object state changes in MIT Plasticine Lab. Planned within the learned latent space to produce actions that reach a desired target shape.
• Encoded visual observations using both DINO/ViT-2 pre-trained and learned embeddings for compact state representation. Trained dynamics model on 50 demonstrations.
• Optimized trajectories with the Cross Entropy Method (CEM), enabling zero-shot generalization to novel target shapes

Custom tendon-driven soft robot for vacuum-based apple harvesting.
• Developed entire data collection pipeline and codebase in ROS2 to capture teleoperated demonstrations
• Trained imitation learning policies like ACT and Diffusion Policy, enabling sample-efficient grasping on only 50 collected episodes
• Designed and tuned position-based PID controllers in firmware for precise tendon control

PennOS: UNIX-like OS implementation in C

From-scratch UNIX-like operating system in C with a multithreaded priority-based scheduler, supporting shell interactions and external storage management via a FAT file system.
*Code not available due to academic integrity policy of CIS 3800 - Operating Systems