A chronicle of past research — most of it in robotic manipulation, grasping under uncertainty, and benchmarks for hand design. Some shipped to conferences, some lived in the lab. All of it shaped what I'm building now.
From prior human studies, we knew humans massively out-grasp robots when objects are in noisy in-hand positions. I took one of the human-adapted strategies as expert demonstration and applied DDPG to teach a robot hand to grasp varied basic objects from randomized in-hand poses. Single-object expert demonstration plus pretraining accelerated the learning pipeline.
Most robotic manipulation benchmarks evaluate either hardware or planning — rarely both. I designed a benchmark that probes both: place the hand at the object center, then perturb the pose outward through increasingly hostile offsets, and measure how many poses the hand still succeeds from. Binary search over poses speeds up trial sequencing; an autonomous test rig runs it without a human in the loop.
Manipulation research usually requires a human to reset the scene between grasps. We built a physical testbed with a robust automated reset mechanism and embedded sensors (including RGB) so a robot can run grasp trials end-to-end. The goal: collect massive real-world data and lower the sim-to-real barrier.
Humans grasp well because most objects are human-centric. Most grasp planners focus on placement, not finger motion — but the moment fingers contact the object often decides success. We coined "near-contact" grasping (hand and fingers very close to the object), ran human studies on multiple robot hands, distilled three primitive control strategies, and turned them into PID controllers. Published & presented at IROS.
Grasp metrics usually rely on physical contact data — which is noisy and hardware-dependent. We collected hundreds of grasp samples, tracked the changing distance between hand and object across each grasp, and trained a classifier to predict grasp quality from geometry alone. Published at ICRA.
Robot hands are designed all over the map — varying in finger count, joint count, actuation, stiffness. We ran human studies on hard manipulation tasks to understand the role each feature plays. I helped run experiments and analyze data; we published at IROS.
While an undergrad research assistant, I built and instrumented the Yale OpenHand Model O — a 4-DOF underactuated gripper tendon-driven from each finger — and wrote the control interface to actuate it. The build that made me a roboticist.
The 4-DOF Barrett Hand struggled with small, thin objects (pens, plates). We designed soft finger pads with tip bumps and stiff "nail" plates that increased contact area. Used SolidWorks, FEA, and 3D printing through the prototyping loop. Presented at OSU's Annual Engineering Expo.