Zhikai (Logan) Zhang

I recently graduated from my MSME-R program in the Department of Mechanical Engineering at Carnegie Mellon University I did my research at the Biorobotics Lab and advised by Professor Howie Choset.

Currently I am working in Biorobotics Lab as a Research Engineer, working on Pipe inspection robot control/prototyping and modular robot learning

I was previously an embedded software intern at Garmin Ltd.

I received my BASc. degree from Department of Eletrical and Computer Engineering, University of British Columbia in 2022.

Email  /  Github  /  Resume

Research/Career Interests

I am deeply passionate about the field of robotics, with a particular focus on modular robots, robot learning, control and planning. My current research is centered around the Modular robot distributed control, reinforcement Learning and planning over control. I hope to explore more on control and planning in robotics, as well as the novel and relatively less explored field of design automation.

Robot Platforms

Modular Robot - Eigenbot

Unitree Go1

CMU ballbot

KUKA IIWA

Robocup Robots - Thunderbots

self made 6DoF robot arm

Research/Projects

Eigenbot walking and running on Terrain in Simulation

Current RL training pipeline gives a centralised end to end policy, which requires heavy tuning and is subjected to unstable performances and local minimum. Such policy needs to be re-trained for different types of robots and is hard to collect enough data, which is lacking in the robotics field. In order to move closer to a robot foundation model, we propose to modularize different components of the robot, eg. legs, vision, arms etc. to be pre-trained on variety of tasks. Such modular networks can thus be combined into different configurations and only needs simpler fine tuning. We hope to simplify the RL design cycle and data collection pipeline with this appraoch. We use reinforcement learning (PPO) to train a walking/running policy for a modular legged robot.

Currently we have Accomplished the following in simulation:
1. successfully trained a centralised RL policy for the hexapod configuration over terrain
2. Used Regularized Online Adaptation (ROA) for Sim2Real transfer of priviliged states
3. distillation of a depth encoder-based policy to use depth image in place of priviliged sandots

Next steps:
1. Testing our policy on Hadware that has action delays and less agile actuators
2. Modularize the policy to be leg/joint specific for locomotion and manipulation
3. Testing a modularized policy ith different configurations (eg. different number of legs, loco-manipulation configurations)

4 leg walking configuration

Hardware walking on flat terrain

Simulation walking on unstable block

Energy Cost Prediction when approaching an obstacle

Sim2Real Work

Real time events are performed with instinct gained from repeated practice, while long term planning optimization requires thinking and understanding of the task at hand. This is analogous to fast and agile movements vs optimal path planning in robotics. Navigating challenging terrains demands a nuanced understanding of the robot's versatile capabilities. Highly agile mobile robots endowed with complex locomotion skills, including climbing, jumping, and walking can often better handle such environments than traditional wheels, but planning for such systems can often be much more difficult. This paper introduces a comprehensive framework for real-time autonomous navigation of such motions by determining path costs using a neural network- based cost predictor by leveraging elevation maps and A* path planning. Additionally, we trained a steerable locomotion policy, where its behaviours are conditioned on both obstacles and waypoint. We demonstrate in simulation that this system can plan paths for multiple different cost functions, such as minimum energy or minimum time, depending on the specified criteria.

Planned Path based on Costr predidtor

Planned Path Execution in Simulation

Ballbot avoid obstacle real time

Ballbot pushing on wall

Several ballbots have been developed, yet only a handful have been equipped with arms to enhance their maneuverability and manipulability. The incorporation of 7- DOF arms to the CMU ballbot has presented challenges in balancing and navigation due to the constantly changing center of mass. This project aims to propose a control strategy that incorporates the arms dynamics. Our approach is to use a simplified whole-body dynamics model, with only the shoulder and elbow joints moving for each arm. This reduces the number of states and accelerates convergence. We focused on two specific tasks: navigation (straight and curved paths) and pushing against a wall. Trajectories were generated using direct collocation for the navigation task and hybrid contact trajectory optimization for pushing the wall. A time-variant linear quadratic regulator (TVLQR) was designed to track the trajectories. The resulting trajectories were tracked with a mean-average error of less than 4 cm, even for the more complex path. These experiments represent an initial step towards unlocking the full potential of ballbots in dynamic and interactive environments.
Supplementary information, including code and animations, can be found at https://github.com/jrapudg/ocrl ballbot navigation project.

KUKA IIWA arm motion planned with RRT-connect

Hololens AR interface developed

Learning from Demonstration (LfD) is a framework that allows lay users to easily program robots. However, the efficiency of robot learning and the robot's ability to generalize to task variations hinges upon the quality and quantity of the provided demonstrations. Our objective is to guide human teachers to furnish more effective demonstrations, thus facilitating efficient robot learning. To achieve this, we propose to use a measure of uncertainty, namely task-related information entropy, as a criterion for suggesting informative demonstration examples to human teachers to improve their teaching skills. In a conducted experiment (N=24), an augmented reality (AR)-based guidance system was employed to train novice users to produce additional demonstrations from areas with the highest entropy within the workspace. These novice users were trained for a few trials to teach the robot a generalizable task using a limited number of demonstrations. Subsequently, the users' performance after training was assessed first on the same task (retention) and then on a novel task (transfer) without guidance. The results indicated a substantial improvement in robot learning efficiency from the teacher's demonstrations, with an improvement of up to 198% observed on the novel task. Furthermore, the proposed approach was compared to a state-of-the-art heuristic rule and found to improve robot learning efficiency by 210% compared to the heuristic rule.

TPGMM online unsupervised Learning

Path Planning

Winning First Place in Robotcup 2019

Control Archetecture Overview

As the Eletrical Sub-team Co-Lead, I spearheaded multiple projects, including Eletrical system redesign, Wifii system integration and FPGA Updates. Technical Work during this Time includes Wifi firmware integration, Robot firmware development, part of Robot Main board design, FPGA intrgration with Wifi system on SPI protocol, Planning Software as well as solenoid curve fitting for better chipping and kicking performance.
Through a Collective Team Effort, we were able to achieve first place in Robocup 2019.

Goal by chipping in Robocup 2019

Dribbler ball steal and shoot for goal with motion Primitive

Robot Arm Version 2

ROS Integration with Robot Arm Version 2

In this personal Project, I have worked with the following:
1. Mechanical, electrical design and Built a robot arm with 3D printed parts & off the shelf motors
2. 6DOF robot dynamics analysis and PD + gravity control, implementation on the physical robot
3. Trjectory generation with 6Dof Arm and control firmware on STM32 platform

Robot arm version 1 with trajectory control, smoother than set point control

Robot arm version 1 with set point control

Experiences

CV based Yaw estimation in noisy reflective pipe

CV based Yaw estimation in field deployment pipe (less noisy)

Pipe Force specializes in intelligent in-pipe surveying and digitizing pipe assets using robotics and AI, providing actionable data that helps infrastructure owners minimize maintenance costs due to pipe failure. This startup stemmed from the Biorobotics Lab
I use Computer Vision + IMU sensor fusion for localisation of the robot's odometery in the pipe environment using Kalman filter. Two pipe environments are tested using this method, as shown in the video.
The estimated states are used in a controller for adjusting the robot's orientation, tested in simulation and harware.

Orienatation Control Simulation vs Hardware

ANT protocl chip module

Lidar-lite V4 wireless sensor

As an Embedded Software Intern, I worked on vrious projects in Garmin, including ANT network (Low energy Wireless network used internally in Garmin), large scale BLE mesh testing and firmware debugging/testing, Lidar-lite V4 sensor, and other frimware testing scripts such as connection tests and battery data analysis.

ANT protocol

BLE Mesh Network