RankQ is an offline-to-online Q-learning objective that augments temporal-difference learning with a self-supervised ranking loss. For VLA RL fine-tuning, RankQ improves low-data regime simulation success by 42.7% and high-data regime simulation success by 13.7% over the next best method. RankQ also enables strong sim-to-real transfer, improving real-world cube stacking success from 43.1% to 88.9% relative to the pretrained imitation baseline.

We leverage 3D world generative models and a language-driven scene designer to scalably generate diverse interactive environments for RL fine-tuning of VLA models. RL fine-tuning across hundreds of generated scenes improves simulation success from 9.7% to 79.8% and real-world success from 21.7% to 75%. We further show that increasing scene diversity substantially improves zero-shot generalization.

Implicit soft-body simulation for scalable soft robot policy learning. Parallel stepping yields up to 6× speedup in non-contact and 40× in contact-rich scenarios over existing frameworks. Cross-simulator evaluations reveal minimal sim-to-sim gap, demonstrating substantial acceleration without loss of physical fidelity.

End-to-end visuomotor policy learning for sim-to-real quadruped mobile manipulation. Policies are trained using a teacher-student framework on a multi-stage pick-and-place task and transfer with ~80% success to real hardware. Emergent behaviors include re-grasping and task chaining, with extensive ablations identifying key ingredients for successful sim-to-real transfer.

Slice-based active learning approach for 3D segmentation. Contrastive learning with Coreset enables metric learning by leveraging inherent data groupings in medical imaging. Evaluations on four datasets reveal enhanced performance over existing active learning methods, achieving cost-efficient low-annotation training.

End-to-end pipeline for full sim-to-real robotic DLO deployment. Physical simulations are used to train a neural controller capable of deploying rods along a prescribed pattern. Scaling analysis is used to obtain a generalizable control policy with respect to material and geometric parameters.

End-to-end pipeline for full sim-to-real robotic paper folding. Neural force manifolds are learned from physical simulation and are used to generate optimal folding trajectories. Scaling analysis is used to obtain a generalizable control policy with respect to material and geometric parameters.

Full scale generalizable physical simulator capable of simulating soft physics, frictional contact, and continuum control. A fully implicit formulation allows for increased computational efficiency while maintaining physical accuracy when compared to previous SOTA.

Robust and rapid instance segmentation of DLOs via skeleton pixel traversals. Segmentations are generated on topologically corrected skeleton structures with intersections handled by the simple and physically insightful optimization objective of minimizing cumulative bending energy.

Fully implicit frictional contact method for simulation of slender elastic rods. Enforces non-penetration and capable of reproducing sticking and sliding phenomena. Case study for flagella bundling in viscous fluid shown.

Study of the snap buckling process of overhand knots. Discrete differential geometry based simulations and tabletop experiments are used to explore the onset of buckling as a function of topology, geometry, and friction.

Preemptive motion planning via human object placement prediction using human pose and gaze as input. Rapidly outperforms traditional wait-and-react methods for human-to-robot pick-and-place operations.

Weak supervision approaches for automated vessel segmentation. Utilizes active contour models to generate weak labels and introduces low-cost human-in-the-loop strategies that drastically reduce labelling cost while maintaining segmentation accuracy.

An implicit contact model for slender elastic rod simulation capable of enforcing non-penetration via smooth penalty energy and edge-to-edge minimum distance formulation. Sliding friction is handled semi-explicitly. Extensive physical validation through knot tying case study.

Implementation of the paper "Hand-Eye Calibration of SCARA Robots" by M. Ulrich. Completed as part of an internship at Vecna Robotics.

ROS wrapper for OpenPose with support for key commercial cameras used by the robotics community.

An Arduino controlled robot dog that I created while healing from a knee surgery. Can do simple functionalities such as sit and respond to praise. So far, the YouTube video has amassed over 190000 views!

A manipulator capable of tying shoes. Built under the design constraints of using at most 2 motors and a $600 budget. Won a robotics competition against Meijo University, resulting in an all expense paid trip to Japan to present the robot. Was featured on several news outlets. So far, the YouTube video has amassed over 59000 views!