Kamal Carter

I am a senior prototype engineer at Boston Dynamics, where my work focuses on the intersection of mechanical design, robot locomotion, software reinforcement learning and controls, and computer vision for the Spot robot.

Previously, I was an application engineer at HEBI Robotics focused on manipulation software, where I worked on applications like series elastic manipulation for door opening, nondestructive material testing, and teleoperation.

I earned my Master's in Mechanical Engineering - Research from Carnegie Mellon University, advised by Aaron Johnson and Sarah Bergbreiter. My graduate research, mentored by Justin Yim, focused on robot locomotion, specifically on developing a scalable, minimally actuated bipedal walker. I also explored Gazebo walking simulations and motion capture control strategies.

Outside of my work, I enjoy playing ultimate frisbee and basketball, and I have a passion for viral robotics side projects.

Email  /  GitHub  /  LinkedIn

Research

I'm interested in robotic athletic intelligence, robot manipulation, and reinforcement learning for whole body control, mechanical design, and computer vision.

Scalable minimally actuated leg extension bipedal walker based on 3D passive dynamics

We present simplified 2D dynamic models of the 3D, passive dynamic inspired walking gait of a physical quasi-passive walking robot. Quasi-passive walkers are robots that integrate passive walking principles and some form of actuation.

Work Projects

Sampling of work projects I've been a part of at HEBI Robotics and Boston Dynamics

[Boston Dynamics] - Spot Quadruped GPS

Mechanical and electrical hardware integration getting Spot to work with GPS for outdoor autonomous navigation.

[Boston Dynamics] - Robot Choreography with Heavy Mass

Worked on the mechanical design of testing how a quadruped robot can do dance routines with up to 50 pound payload. Worked on a prototype learned controller in mujoco that optimize for energy efficiency for saving battery life, never deployed but fun.

[HEBI ROBOTICS] - 4DOF Arm Robot Scanning

For non desctrutive testing worked on the teleoperation of using a 4 DOF to do ultrasonic scanning thickness of metal areas. Each actuator had torque sensing to at the endeffector tip axis had robot provide a constant output force to get consistent readings from the tool tip

[HEBI ROBOTICS] - Leader Follower Robot

Remap kinematics of a 7DOF robot encoder arm to 7 DOF actuated robot arm for customers to do non destrutive testing without prolong exposure to dirty enviornments.

[HEBI ROBOTICS] - Computer Vision Door Opening

Worked on the mapping controller for kinematics of a 6DOF robot arm on an articulating tracked base. Computer Vision using Yolo to detect door handles and create door opening pathing based on handle location to turn the door.

[HEBI ROBOTICS] - Track Robot Navigating Cluttered Enviornment

Early work was on the teleoperation of a 8DOF articulating track robot. Intially user had individual control of each track base on their tablet controller. Later started working on using joint torque feedback from the actuators to have the robot autonomously provide consistent torque to the ground for better navigation of stair cases and cluttered enviornments.

Other Projects

These include coursework, side projects and unpublished research work.

Humanoid Reinforcement Learning Track and Field [IN PROGRESS]

Saw the Humanoid Olympic event in China and decided to instead of teleoperating robot have the robot do the track events autonomously. Using Unitree’s H1 Humanoid to learn different track events (40 yard, 100 meters, 110 meters hurdle, 200 meters, 400 meters, 1500 meters) using RL reward policy. In the future I want to do retarting work using visual data gather especially for the 40 yard dash since there is a lot of training footage from YouTube as a way not to have to spend so much time tuning intial rewards. Also in the future I want to add visual perception of the white lines on the track so instead of following the path I set can use visual feedback.

Robot Aimbot Valorant

Used fast object detection YOlO optimized to extremely low latency 100 Hz with a mechanical moving table to react to video game targets on the screen

Fast Quadruped Crawling - Reinforcement Learning on Inexpensive Hardware

Built a 8Dof quadruped robot that uses reinforcement learning for walking. The robot is trained in Pybullet Simulator to maximize robot travel speed contrained by actuator specifications and limits. The robot position based trajectory is then transfered from the simulation on real world hardware

Monkey Robot - Triple Pendulum

Developed a bioinspired, three-link swinging robot, in the vein of a triple pendulum, that can modify its motion by means of motor actuation of its lower two links (its tail). By comparing the three-link robot’s ability to swing from rest and reorient in-air to its two-link counterpart, we can study more generally how tails can modify robotic movement, and theoretically gain a better understanding of robotic locomotion. A MATLAB trajectory optimization simulation of the two and three-link robot indicated that a three-link configuration increased maximum achievable velocity while decreasing required torque when swinging from rest.

Mini Ballbot

Work with Roberto Shu in CMU RI course 16-264 Humanoids, built a low cost robot that balances on a ball. Miniballbot was controled using full state feedback with PID control.

Design and source code from Jon Barron's website