# Research

This page lists my accepted/in-review publications with more detail, links to the papers, and supplemental material like media, code, and citations. I also provide short descriptions of my current research, but withhold some of the important technical details since the work is ongoing. Click on project titles to see more detail. For more information on my current research, please email me directly. For a more condensed overview of my research, look at my CV.

## Journal Publications

Tensegrities are structures traditionally comprised of rigid bars held in equilibrium by a network of cables. Originally explored by architects and civil engineers, it was not until relatively recently that actuated tensegrities were studied in the field of robotics. As robots, tensegrities enjoy benefits such as lightweight construction and adjustable compliance, exhibiting features of both fully rigid and fully soft robots.

This project studied what we term compound tensegrities, whose rigid elements are not just purely compressive bars, but can take on any arbitrary geometry (and therefore also admit net moments about each rigid body). This allows more faithful bio-inspired designs at the expense of complicating traditional tensegrity modeling techniques. In our work, we reformulate the classic force density method used for computing equilibrium forces in network structures to allow for analysis of compound tensegrities and derive a fast optimization-based shape controller that can also be used for quasi-static position control. Finally, the method is validated both in simulation and on a simple hardware test case.

Andrew Preston Sabelhaus, Albert Hao Li, Kimberley Sover, Jacob Madden, Andrew Barkan, Adrian Agogino, and Alice Agogino, "Inverse Statics Optimization for Compound Tensegrity Robots," IEEE Robotics and Automation Letters, vol. 5, no. 3, pp. 3982-3989, 2020.

@article{
Sabelhaus2020_invstatopt,
title={Inverse Statics Optimization for Compound Tensegrity Robots},
author={Sabelhaus, Andrew Preston and Li, Albert Hao and Sover, Kimberley and Madden, Jacob and Barkan, Andrew and Agogino, Adrian and Agogino, Alice},
journal={IEEE Robotics and Automation Letters},
volume={5},
number={3},
year={2020},
pages={3982-3989},
}


## Conference Publications

Humans are very skilled at spatiotemporal prediction, being able to predict the motion of objects or other agents with fairly minimal observations. This project sought to learn latent dynamics models for the purpose of long-horizon spatiotemporal prediction on robots. As a prerequisite, we must be able to extract latent states from sequences of observations. For example, we may be interested in predicting positions of objects (the states) from video frames (the observations). This process is called posterior inference.

Many conventional methods in the learning literature conduct posterior inference using an inference network like a bi-RNN, which consumes the data and returns an estimate over some latent state. As it turns out, this is unnecessary, as classical state estimation theory allows approximate inference using methods like the extended Kalman filter, which requires only a dynamics network and an observation model, eliminating the need for a third inference network.

We present results for such a model and also a new learning algorithm called replay overshooting that prioritizes training the dynamics model over the observation model. We show that our method is effective on multiple types of data, such as stripped position data or image sequences, while remaining very parameter-efficient compared to existing methods in the literature. As an added bonus, our method allows seamlessly changing the learned model between discrete and continuous-time, the first such method that natively incorporates both perspectives for dynamics learning.

Albert Hao Li*, Philipp Wu*, Monroe Kennedy III, "Replay Overshooting: Learning Stochastic Latent Dynamics with the Extended Kalman Filter," 2021 IEEE International Conference on Robotics and Automation (ICRA), Xi'an, China, 2021, pp. 852-858. *Equal Contribution.

@inproceedings{
liwu_2021ReplayOvershooting,
title={Replay Overshooting: Learning Stochastic Latent Dynamics with the Extended Kalman Filter},
author={A. {Li} and P. {Wu} and M. {Kennedy}},
booktitle={2021 International Conference on Robotics and Automation (ICRA)},
year={2021},
pages={852-858},
}


To facilitate human-robot interactions and robotic multi-tasking in dynamical environments, we expect them to be able to reason accurately about contact. One of the simplest benchmarks for controlling via contact is the robotic juggling task, which has been replicated before on quadrotors, robotic manipulators, and large humanoid robots.

This project studied bounce juggling on the bipedal robot Cassie, developed by Agility Robotics. The distinguishing factor of this work is the difficulty of maintaining an active juggle while also balancing the robot, which is already a nontrivial task depending on the robot's environment. We derived two different juggling controllers, both inspired by mirror law algorithms: one based on a simple PID scheme on the ball position and another based on contact force optimization at the robot's feet to maintain balance while driving the juggling with sufficient force and precision.

We found that in simulation, both controllers were locally exponentially stable with Poincaré analysis. We implemented the simpler of the two on Cassie and found that we were able to maintain stable juggles for long periods (with the longest period being over 40 juggles long). Unfortunately, I left Berkeley to start my MS near the end of this project, and we did not have time to implement the second controller online.

Katherine Lin Poggensee*, Albert Hao Li*, Daniel Sotsaikich*, Bike Zhang, Prasanth Kotaru, Mark Mueller, and Koushil Sreenath, "Ball Juggling on the Bipedal Robot Cassie," 2020 European Control Conference (ECC), Saint Petersburg, Russia, 2020, pp. 875-880. *Equal Contribution.

@inproceedings{
Poggensee2020_jugglingcassie,
title={Ball Juggling on the Bipedal Robot Cassie},
author={Poggensee, Katherine Lin and Li, Albert Hao and Sotsaikich, Daniel and Zhang, Bike and Kotaru, Prasanth, and Mueller, Mark and Sreenath, Koushil},
journal={2020 IEEE European Control Conference (ECC)},
year={2020},
pages={875-880},
}


Vacuum-based grasping and manipulation is a popular alternative to traditional parallel-jaw methods, since we only need to compute a single point of contact and we may be able to lift objects that are difficult to grasp otherwise. The goal of this project was deriving an analytical physics-based suction model, training a grasping planner to find good suction points on various objects from point cloud data, and implementing the system on a real robot.

My role on the project was designing new grippers that could interface with the ABB YuMi and replace its default grippers with lightweight and cost-effective alternatives. I was also responsible for hardware maintenance, including part replacements, fixing electronic parts and circuitry-related failures, and implementing new design features. Ultimately, Dex-Net 3.0 was able to achieve success rates of 98%, 82%, and 58% on the "basic," "typical," and "adversarial" object categories, with adversarial performance improving to 81% when trained specifically on adversarial objects.

Jeffrey Mahler, Matthew Matl, Xinyi Liu, Albert Li, David Gealy, Ken Goldberg, "Dex-Net 3.0: Computing Robust Vacuum Suction Grasp Targets in Point Clouds Using a New Analytic Model and Deep Learning," 2018 IEEE International Conference on Robotics and Automation (ICRA), Brisbane, QLD, 2018, pp. 5620-5627.

@inproceedings{
Mahler2018_dexnet3,
title={Dex-Net 3.0: Computing Robust Vacuum Suction Grasp Targets in Point Clouds Using a New Analytic Model and Deep Learning},
author={Mahler, Jeffrey and Matl, Matthew and Liu, Xinyu and Li, Albert and Gealy, David and Goldberg, Ken},
journal={2018 IEEE International Conference on Robotics and Automation (ICRA)},
year={2018},