GoodLabs Robbie: An Advanced Humanoid for Geriatric Care
The group of people aged 80 years or older, the oldest old, is the fastest-growing age segment. By 2050, this segment is projected to triple, reaching 426 million. The rising “oldest old” population urgently requires mobility solutions to ensure their physical well-being, dignity, and autonomy. Mobility directly impacts seniors’ self-worth and mental health. Addressing these challenges bolsters their health and reinforces their valued role in our communities.
GoodLabs Studio, an extreme A.I. engineering studio founded with a mantra of creating a better tomorrow, is developing a bipedal robot Robbie to eventually assist seniors’ mobility and house chores as a lifelong companion.
Robbie Mark-I is a humanoid robot designed to achieve natural walking and eventually serve as a foundation to aid in eldercare.
Over the past few months, GoodLabs’ Team Robbie has designed and built our first prototype, a 3-foot-tall Robbie Mark-I, and a reinforcement learning algorithm to train Robbie to walk. This series of blog posts document the entire uncensored journey of how we develop Robbie Mark-I, including all the failures we have encountered, to provide a glimpse of how difficult it is to create a sophisticated bipedal robot, let alone for eldercare.
Robbie Mark-I is equipped with the following:
An Nvidia Jetson Nano
An Arduino Mega 2560
21 motors with torque up to 80kg/cm
2 7.4V Li-Po 2-cell batteries
Voltage regulator and stepdown boards
I2C to PWM signal conversion boards
2 Gyroscope Accelerometer boards
Various sensors (Positional potentiometers and temperature sensors)
Over 50 custom designed 3D printed parts
All parts are printed using three high-speed 3D printers to reduce build time. We initially used simple PID loops to run the balancing and locomotion algorithm. However, we discovered they were only effective for a small range of environments. Therefore, we decided to use a Jetson for the prototype to run a more adaptive machine learning system. At over 25 times a second, the onboard computer will send data packets of the system outputs to a microcontroller which executes smoothing using PID loops and ultimately dictates the movement of the distributed motors, thereby controlling the robot.
The AI system we designed specifically for humanoid robots works off a network of policies, the most crucial being the adversarial reinforcement learning model. This model takes suggestions for expected pose trajectories from a trajectory estimation model, which are run under Issac Gym and judged under two evaluation networks for pose realism and transition realism.
In the next article, we will discuss more in-depth how we use Isaac Gym to Robbie Mark-I's reinforcement learning model. Stay tuned!