Albert Wu

I am a 5th year Ph.D. student at the Computer Science Department of Stanford University. I study robotic manipulation at The Movement Lab. My advisor is C. Karen Liu.

I received my B.S. and M.Eng. from MIT EECS advised by Russ Tedrake. I also worked as an undergraduate researcher in Hugh Herr's lab.

Email  /  GitHub  /  Google Scholar  /  YouTube

Research

My research interst includes robotic manipulation, motion planning, and control. I seek to develop robust and data-efficient motion planners for contact-rich systems through fusing physical models and learning methods.

In the Wild Ungraspable Object Picking with Bimanual Nonprehensile Manipulation

We developed a method to retrieve grocery items that are difficult to grasp using traditional suction and parallel grippers. Through extensive testing in a replica store and a real-world grocery store, our mobile manipulator demonstrated high success rates in retrieving diverse ungraspable items from cluttered shelves.

One-Shot Transfer of Long-Horizon Extrinsic Manipulation Through Contact Retargeting

The use of environment contacts enables manipulation strategies that are otherwise impossible with a parallel jaw gripper. We propose to generalize one extrinsic manipulation trajectory to diverse objects and environments by retargeting contact requirements. Using a 7+1 DoF robotic arm-gripper system, we achieved an overall success rate of 80.5% on hardware over 4 long-horizon extrinsic manipulation tasks.

Synthesize Dexterous Nonprehensile Pregrasp for Ungraspable Objects

We combine graph search, optimal control, and a learning-based objective function to synthesize dexterous pregrasp sequences for objects in initially ungraspable.

Learning diverse and physically feasible dexterous grasps with generative model and bilevel optimization

We use a generative model and a bilevel optimization to plan diverse grasp configurations on novel objects for a 4-fingered robotic hand. Our method achieved 86.7% real-world success rate on 20 household objects with unseen and challenging shape.

Real-time model predictive control and system identification using differentiable simulation

We present a method for continuous improvement of modeling and control after deploying the robot to a dynamically-changing target environment. We develop a differentiable physics simulation framework that simultaneously performs online system identification and optimal control using the incoming observations from the target environment in real time. Our method outperforms baseline methods on simulation and hardware experiments.

Robust-RRT: Probabilistically-complete motion planning for uncertain nonlinear systems

We propose Robust-RRT, which integrates forward reachability analysis with rapidly-exploring random tree. Unlike exisiting robust planning algorithms, Robust-RRT is theoretically sound without restricitng the system structure. Specifically, Robust-RRT is probabilistically complete for nonlinear Lipschitz continuous dynamical systems with bounded uncertainty. Using sampling-based reachability analysis, we demonstrate Robust-RRT on simulated nonlinear, underactuated, and hybrid systems.

The Nearest Polytope Problem: Algorithms and Application to Controlling Hybrid Systems

Given a list of polytopes and a distance metric in Euclidean space, we seek an empirically fast algorithm to find the nearest polytope to a query point. We discuss the properties of 3 proposed algorithms and compare their performances using datasets motivated by control applications, including sampling-based motion planning and model predictive control.

R3T: Rapidly-exploring random reachable set tree for optimal kinodynamic planning of nonlinear hybrid systems

R3T addresses the difficulty of defining distance in kinodynamic and hybrid systems by performing tree expansion with the (approximated) forward reachable sets. Under mild assumptions, R3T is probabilistically complete and asymptotically optimal through rewiring. The advantages of R3T are demonstrated on simulated nonlinear, hybrid, and contact-rich robotic systems.

An autonomous exoskeleton for ankle plantarflexion assistance

We present an untethered, autonomous exoskeleton platform for ankle plantarflexion assistance. The exoskeleton is driven by bowden cables, which allows the actuator to be placed at the hip for reduced distal mass and inertia.

Other Projects

These include coursework, side projects and unpublished research work.

Changing Utensils with Reinforcement Learning

I applied reinforcement learning to Dexai Robotics’s food serving robot for autonmous utensil change.

Inertial Wheel Pendulum

I made an intertial wheel pendulum for the MIT course 6.115 Microcomputer Project Lab. The nonlinear energy-shaping controller and local LQR controller are from an old assignment problem in 6.832 Underactuated Robotics.

Website design and source code from Leonid Keselman and Jon Barron.