Projects

Playful Quadrupeds

My work explores how large language models (LLMs) and reinforcement learning from human feedback (RLHF) can be used to make quadrupeds like Pupper more friendly and expressive. I previously worked on deep reinforcement learning to enable Pupper to walk faster on more challenging terrains.
I work for the nonprofit Hands-On Robotics, and we are collaborating with the Stanford Children’s Hospital to explore how quadrupeds can improve peri-operative and inpatient pediatric experience. Our work has been featured in the Stanford Report, ABC News, and more.

AI Robotics Education

In 2022 I co-taught and designed an introductory AI Robotics course to undergraduates at Stanford. In 2023, the course was elevated to CS 123 and I served as head TA for Professor Karen Liu.

Game Theoretic Planning

In the Stanford Multi-Robot Systems Lab, I am exploring how learned objectives can improve speed and precision of game-theoretic planners for automous vehicles. We published our most recent work here.

Previous: Online Re-planning for UAVs

Multiple Unpiloted Aerial Vehicles (UAVs) working together have the potential to efficiently survey large geographical areas. Unfortunately, UAVs in the field may fail midway through a survey due to adverse weather, faster-than-expected battery drain, or mechanical malfunction, leaving part of the survey area uncovered. Here we propose an algorithm to online re-plan coverage routes for multiple UAVs to take over the remaining route of a failed team member. We first present a greedy path recovery algorithm whereby each UAV greedily absorbs the closest remaining vertices from the failed UAV’s route into its own route. This method is then extended using an existing Tabu search method for multi-agent path repair to give successively better quality paths. We call the new path repair algorithm GRIT (Greedy Repair Initializes Tabu search), and demonstrate it performing path repair for nominal paths planned with both a traditional lawnmower-style planner and a more sophisticated integer program based planner. We show that GRIT achieves adequate re-plans 10-50 times faster than two benchmark planners, making it ideal for online path repair in mid-flight, although the benchmarks eventually outperform GRIT if given unlimited computation time.

Previous: Characterizing Key White Shark Habitat

White sharks are known to migrate across the Pacific Ocean to the lee of Hawai’i Island in the spring and early summer. Little is known about why white sharks perform this migration and how key oceanographic features like eddies and seamounts affect their behavior. We analyzed novel white shark Pop-up Satellite Tag (PSAT) data tracking both males and females migrating to the lee of Hawaii and performing diel vertical migrations. We also performed in-situ sampling of oceanography in the region, targeting eddies, eddy fronts, and seamounts to rationalize these migrations. Data from CTD hydrocasts, ADCP sampling, and meter net tows suggest that the core of anticyclonic eddies, peripheries of cyclonic eddies, and seamounts are structuring vibrant mesopelagic communities in the area.

Previous: Gliding Lizard Flight Dynamics

The ability to glide through an arboreal habitat has been acquired by several mammals, amphibians, snakes, lizards, and even invertebrates. Lizards of the genus Draco possess specialized morphological structures for gliding, including a patagium, throat lappets, and modified hindlimbs. Despite being among the most specialized reptilian gliders, it is currently unknown how Draco is able to maneuver effectively during flight. Here, we present a new computational method for characterizing the role of tail control on Draco glide distance and stability. We first modeled Draco flight dynamics as a function of gravitational, lift, and drag forces. Lift and drag estimates were derived from wind tunnel experiments of 3D printed models based on photos of Draco during gliding. Initial modeling leveraged the known mass and planar surface area of the Draco to estimate lift and drag coefficients. We developed a simplified, 3D simulation for Draco gliding, calculating longitudinal and lateral position and a pitch angle of the lizard with respect to a cartesian coordinate frame. We used PID control to model the lizards’ tail adjustment to maintain an angle of attack. Our model suggests an active tail improves both glide distance and stability in Draco. These results provide insight toward the biomechanics of Draco; however, future in vivo studies are needed to provide a complete picture for gliding mechanics of this genus. Our approach enables the replication and modification of existing gliders to better understand their performance and mechanics. This can be applied to extinct species, but also as a way of exploring the biomimetic potential of different morphological features. See the paper.

Jaden Clark