Annan Zhang

PhD Student
Robotics & Artificial Intelligence
Massachusetts Institute of Technology (MIT)

Email  /  Google Scholar  /  LinkedIn  /  GitHub

I am a PhD student at the MIT Computer Science & Artificial Intelligence Lab (CSAIL), advised by Daniela Rus. My research interests lie at the intersection of robotics, materials, and machine learning. My goal is to combine compliance and machine learning to develop physically intelligent robots.

Previously, I earned a master's degree in electrical engineering and computer science from MIT and a master's and a bachelor's degree in mechanical engineering from ETH Zürich. At ETH, I worked with Dirk Mohr on machine learning-based modeling of mechanical material behavior.

I spent a summer working at the asset management company Vanguard, where I explored the use of artificial intelligence and large language models in finance.

In my free time, I enjoy reading and traveling.

(* indicates equal contribution)

We introduce the Neural Jacobian Field, an architecture that learns to control robots by observing the execution of random commands. Our method makes no assumptions about the robots' materials, actuation, or sensing, and enables self-modeling and visuomotor control of unknown robots.

We present a robotic gripper with integrated sensing made from 3D printed elastomer lattices with embedded air channels. Our method simplifies the fabrication process for sensorized grippers and provides sufficient sensor resolutions to reason about grasp location and grasp forces.

We present a grocery packing robot that can pack a stream of unknown objects on a conveyor belt. By integrating multiple sensing modalities, our system estimates object size and stiffness to avoid damaging packings.

Based on experiments on two large soft robotics datasets, we derive best practices for training neural networks that map sensor signals to soft robot shape.

We present a method to generate large collision-free regions in configuration space for sampling- and optimization-based motion planning. We extend the theoretical framework in Amice et al. (2022) and generalize the algorithm to handle algebraic joints and non-polytopic geometries.

We use cameras to record the interior of compliant electric actuators and train a neural network that maps the visual feedback to the actuator's tip pose. Our method presents a robust approach for sensorizing hollow-bodied actuators and provides accurate predictions in the presence of external disturbances.

We embed air-filled channels within architected materials and measure the pressure change during deformation. Our method integrates programmed mechanical behavior, sensing, and actuation and enables sensorized structures for wearables and robotics from one single build material.

We use convex optimization to generate regions in configuration space that are guaranteed to be collision-free. Our method scales to high-dimensional robot manipulators and paves the way for motion planning with verifiable non-collision.

We present a pipeline to simulate and fabricate cable-driven soft robots with embedded skeletons. These hybrid soft-rigid robots combine the best of both worlds and simultaneously provide strength and robustness.

We demonstrate how a plasticity model widely used for ductile engineering materials can be learned by a neural network. We deploy the neural network in commercial finite element software and show its capability to correctly predict stresses from strains.

• Best Paper Award at WAFR 2022, the premier conference on algorithms for robotics
• ETH Excellence Scholarship awarded by the Rector of ETH for my visions in the Materials 4.0 age
• ETH Outstanding Bachelor Award for graduating as valedictorian of the mechanical engineering class of 2019
• ETH Outstanding Bachelor Award for finishing 2nd out of 543 in the selective first-year exams

Last updated: August 2024

Template