> INITIALIZING...
> LOADING PROFILE...
> READY
Embodied AI · Robot Learning · Safe Autonomy
> I build learning-enabled robotic systems that perceive, plan, and act safely in the physical world.
> Robotics Engineering Intern @ AgRobotics
> Master's in Autonomy & Robotics @ UIUC, graduating Dec 2026
End-to-end autonomy for an agricultural robot at AgRobotics: a fine-tuned pi0.7 VLA policy for harvesting, with dual-arm planning, berry perception, and in-field navigation.
Real-time impedance force control for a KUKA LBR Med 7 + hexapod, part of an ARPA-H-funded AI-guided microsurgery project at UIUC.
A morphology-agnostic optimizer that turns unsafe VLA-generated trajectories into collision-free, constraint-satisfying motion.
An end-to-end pipeline turning unseen YouTube cooking videos into executable robot commands, trained on 1.6k+ clips.
A cross-attention ViT conditioned on kinematic features, producing representations that transfer to unseen humanoid robots.
ZMP preview control, A* footstep planning, and MPC balance optimization for the Unitree G1, with RL-based locomotion.
Where I currently go deepest:
Everything I've built, spanning robot learning, manipulation, autonomous vehicles, and embedded systems.
A whole-body stack using ZMP preview control, A* footstep planning, and MPC balance optimization, integrated with RL-based locomotion and Jacobian IK for coordinated walk-and-reach tasks on the Unitree G1.
Problem: standard PD control for humanoid balance is energy-inefficient and struggles to recover from external perturbation.
My contribution: implemented the ZMP preview controller, A* footstep planner, and MPC balance optimization, then integrated RL-based locomotion with Jacobian IK for walk-and-reach tasks.
Limitation: validated in MuJoCo simulation only; not yet tested on physical hardware.
A morphology-agnostic optimizer that turns unsafe VLA-generated trajectories into collision-free, constraint-satisfying motion across 5 robot morphologies (6-23 DOF), in 0.03-0.44s per trajectory.
Problem: VLA policies can output trajectories that violate joint limits or collide with the environment, and those failure modes don't transfer cleanly across robots with different kinematics.
My contribution: Built the full optimization layer, including a fast hierarchical approach (ProxQP + CHOMP) and a slower coupled-dynamics approach using Pinocchio RNEA.
Trade-off: the coupled-dynamics formulation gets better energy results but is meaningfully slower than the hierarchical approach, so the right choice depends on the latency budget.
Imitation learning from pixels on a MuJoCo Franka Panda: a CNN Diffusion Policy trained on 200 scripted-expert demos from RGB (wrist + front) and proprioception only, benchmarked against an ACT baseline and exported to ONNX.
Problem: the v1 policy hit 87% nominally but couldn't recover from a missed grasp and dropped to 67% under distractors and lighting changes.
My contribution: diagnosed the failure modes, added 100 recovery demos from a perturbation expert (with the mistake actions masked out of training) plus train-time distractors and photometric augmentation, and ran ablations with paired McNemar tests on identical seeds.
Finding: recovery data was the biggest win on its own; augmentation only helped on top of it, and alone it cut nominal success from 87% to 57%.
MuJoCo running fully in the browser: a Unitree G1 you can shove while a capture-point balance controller keeps it standing (250N forward push vs. 50N for bare servos), and a Franka pick-and-place driven live by the diffusion-pick ONNX policy.
An end-to-end pipeline that extracts actions from unconstrained cooking videos and maps them to robot primitives. Trained a 19M-parameter temporal model on 1.6k+ clips, including a novel Indian-cooking dataset.
Problem: teaching a robot new manipulation skills usually means collecting robot-specific demonstration data, which doesn't scale to new tasks.
My contribution: built the pipeline end-to-end, including the YouTube action-extraction step, the 19M-parameter temporal model, and the Indian-cooking dataset used to test generalization.
Limitation: accuracy drops on cuisines and actions underrepresented in training (60-70% zero-shot transfer to Western cuisine vs. 77-85% overall), so it's not yet robust to fully out-of-distribution tasks.
Intent-based navigation for GEM e4 using GPT-4o for goal localization, BEV perception, and Stanley control. 88% accuracy on KITTI for language-to-goal localization.
Zero-shot 6D pose pipeline combining VLM semantic inventory, SAM-3 segmentation, and Objaverse mesh retrieval. 76.5-100% ADD-S AUC on YCB-Video, plus an RGB-D variant with ensemble mesh selection.
Real-time pedestrian safety stack fusing LiDAR clustering and RGB-D detection, with Time-to-Collision estimation for proactive speed adaptation. 90% success across tested pedestrian scenarios.
A cross-attention ViT conditioned on 32-dim kinematic features, producing embodiment-aware representations that generalize to unseen humanoid robots with morphologically meaningful embedding structure.
Problem: visual representations trained around one robot's embodiment usually don't transfer to a robot with a different kinematic structure.
My contribution: designed and trained the cross-attention ViT and the 32-dim kinematic conditioning scheme.
Limitation: behavioral transfer, actual task success rates across robots, remains unproven due to rendering-environment constraints; the current evidence is representation quality, not closed-loop task performance.
Vision-based handover with real-time MediaPipe hand tracking, Jacobian IK, velocity limiting, and command smoothing for stable object transport and adaptive release.
Terrain-adaptive locomotion trained with PPO, filtered by a real-time CBF safety layer. Zero-fall locomotion, 99% unsafe-action rejection, 90% of baseline traversal speed retained.
Perception pipeline converting depth clouds to elevation maps for quadruped footstep placement.
ANYmal quadruped policies trained via Proximal Policy Optimization, with robust traversal on irregular terrain under zero-fall constraints.
Fused Visual-Inertial Odometry with footstep planning to enable navigation in GPS-denied environments with <10cm drift.
A 5-DOF prosthetic arm controlled via EMG signals, with real-time gesture classification on-device at 96% accuracy.
Robots that learn tasks from watching humans, then execute them safely in the real world.
> # WHAT I'M EXPLORING
How learned behaviors can transfer across robot morphologies while staying physically feasible and safe to execute, combining cross-embodiment perception with trajectory optimization safety layers.
> # ABOUT_ME
I'm a robotics engineer working across robot learning, perception, planning, and control. I like building systems that connect modern AI with the physical realities of robotics, from learned representations and VLA models to real-time force control on actual hardware.
Always up for talking robotics, embodied AI, and hard autonomy problems.