Hosik Chae

BALLU: An affordable and safe-to-interact-with, but highly underactuated bipedal robot that balances at all times, supported by a set of buoyant objects
- Created the first walking controller using data-driven methods (statistical metrics, imitation learning, and reinforcement learning) to handle severe underactuation, aerodynamic effects, and low-fidelity hardware; validated on hardware with stable limit-cycle behavior
- Trained reinforcement learning (RL) policies for diverse… Show more

BALLU: An affordable and safe-to-interact-with, but highly underactuated bipedal robot that balances at all times, supported by a set of buoyant objects
- Created the first walking controller using data-driven methods (statistical metrics, imitation learning, and reinforcement learning) to handle severe underactuation, aerodynamic effects, and low-fidelity hardware; validated on hardware with stable limit-cycle behavior
- Trained reinforcement learning (RL) policies for diverse contact-rich tasks including walking and turning, achieving over 4× reduction in trajectory tracking error in real hardware tests compared to a simulation-only baseline
- Enabled zero-shot sim-to-real policy transfer by integrating RL-based residual dynamics modeling, analytic modeling, system identification, and uncertainty modeling for low-fidelity sensors
- Established a hybrid control architecture: on-board testing and off-board training via wireless communication to accommodate limited compute and data logging constraints
- Reduced microcontroller CPU load by 50% through hardware-software co-optimization of the communication stack
- Implemented custom simulation modules to supplement unstable or unsupported physics (e.g., aerodynamics, linkage kinematics) for improved numerical robustness and speed Show less

LIMMS: Multi-modal/modular delivery robots that can work as manipulators, vehicles, or legs when attached to a package box.
- Developed a dynamic manipulation controller for a nonprehensile box-catching task with nonlinear optimization
- Guided the team in developing the control architecture and motion planners for manipulator and quadruped modes

Robocup: Fully autonomous soccer competition with two adult-size humanoids as a team.
- Led the high-level software team, developing recognition, localization, and path planning modules, and integrated them with low-level controllers via a ROS and shared memory-based architecture
- Developed global and local path planners through comparative analysis of heuristic, sampling-based, graph-based, and model predictive controller (MPC) approaches
- Optimized a Python-based framework by… Show more

Robocup: Fully autonomous soccer competition with two adult-size humanoids as a team.
- Led the high-level software team, developing recognition, localization, and path planning modules, and integrated them with low-level controllers via a ROS and shared memory-based architecture
- Developed global and local path planners through comparative analysis of heuristic, sampling-based, graph-based, and model predictive controller (MPC) approaches
- Optimized a Python-based framework by applying Ahead-of-Time compilation and rewriting bottleneck modules in C++ with PyBind11, reducing startup time by 80% and CPU usage by 70% Show less

YORI: An anthropomorphic cooking robot with dual 5 DoF manipulators with extra DoF at the torso.
- Developed a task-space impedance controller and a position-force hybrid controller to achieve different primitive motions for cooking tasks under various interactions with the environment
- Implemented a position/force controller for tasks such as scraping a pan with a spatula
- Formulated a task scheduling algorithm with mixed-integer quadratic programming (MIQP) to minimize total… Show more

YORI: An anthropomorphic cooking robot with dual 5 DoF manipulators with extra DoF at the torso.
- Developed a task-space impedance controller and a position-force hybrid controller to achieve different primitive motions for cooking tasks under various interactions with the environment
- Implemented a position/force controller for tasks such as scraping a pan with a spatula
- Formulated a task scheduling algorithm with mixed-integer quadratic programming (MIQP) to minimize total execution time Show less

ALPHRED: A torque/position-controlled 12 DoF multi-modal quadruped that can creep, trot, jump, and perform manipulation tasks.
- Enabled autonomous footstep planning by developing an online vision filter to detect steppable regions from RGBD data, integrated with MIQP-based planners incorporating stability conditions for quadruped locomotion
- Implemented single-arm and dual-arm manipulation controllers for button pressing and box lifting using position and hybrid force-position control… Show more

ALPHRED: A torque/position-controlled 12 DoF multi-modal quadruped that can creep, trot, jump, and perform manipulation tasks.
- Enabled autonomous footstep planning by developing an online vision filter to detect steppable regions from RGBD data, integrated with MIQP-based planners incorporating stability conditions for quadruped locomotion
- Implemented single-arm and dual-arm manipulation controllers for button pressing and box lifting using position and hybrid force-position control, respectively
- Developed a real-time MPC controller for quadruped locomotion, formulating ground reaction force optimization as a quadratic program based on a simplified dynamics model
- Built a shared-memory-based modular framework to support high-performance computing and multiprocessing Show less

- Built custom simulator extensions by deriving analytical models for tendon-driven actuation and aerodynamic effects, stabilizing simulation against divergence and reducing the reality gap to hardware
- Trained RL walking policies for an underactuated tendon-driven biped using a self-implemented Proximal Policy Optimization (PPO) algorithm, enforcing actuator-bandwidth limits and energy objectives to obtain stable, efficient gaits
- Collected motion-capture kinematic datasets during… Show more

- Built custom simulator extensions by deriving analytical models for tendon-driven actuation and aerodynamic effects, stabilizing simulation against divergence and reducing the reality gap to hardware
- Trained RL walking policies for an underactuated tendon-driven biped using a self-implemented Proximal Policy Optimization (PPO) algorithm, enforcing actuator-bandwidth limits and energy objectives to obtain stable, efficient gaits
- Collected motion-capture kinematic datasets during on-hardware rollouts and performed domain adaptation and transfer learning to deploy simulation-trained policies on the real-world robot
- Collaborated within the ML research team to translate hardware constraints into training and adaptation requirements, aligning simulation assumptions with on-robot behavior Show less

SiLVIA: An 18 DoF hexapod developed for general tasks focusing on wall climbing and carrying a heavy payload up to twice its body weight.
- Worked on motion planning and hardware maintenance

NABi: A torque/position-controlled 6 DoF biped robot
- Developed simulation architecture and simulation model

THOR: A disaster relief robot that participated in DARPA Robotics Challenge.
- Worked on improving the ZMP-based walking controller and maintained the hardware

LARA: A humanoid concierge robot and HRI research platform that provides guests with information about the conference center and UCLA’s campus
- Designed software architecture and HRI-related components (body motion, facial expression) and developed facial recognition and tracking module

An actuator generates vortex rings that can transfer its momentum through long distances, which works as an external actuator to move a distant object.
- Analyzed the system based on fluid dynamics, and devised actuator mechanism
- Designed a Preview controller to deal with the discrete/delayed actuation for continuous system