We present the Duke Humanoid, an open-source 10-degrees-of-freedom humanoid, as an extensible platform for locomotion research. The design mimics human physiology, with minimized leg distances and symmetrical body alignment in the frontal plane to maintain static balance with straight knees. We develop a reinforcement learning policy that can be deployed zero-shot on the hardware for velocity-tracking walking tasks. Additionally, to enhance energy efficiency in locomotion, we propose an end-to-end reinforcement learning algorithm that encourages the robot to leverage passive dynamics. Our experiment results show that our passive policy reduces the cost of transport by up to 50% in simulation and 31% in real-world testing. #machine#Robotics#robot
Introducing the Duke Humanoid, an open source 10 degrees of freedom extensible research platform
for humanoid locomotion. Standing one meter tall and weighing 30 kilograms, the Duke Humanoid is
the size of an eight-year-old child. Its leg proportions replicate that of a human leg,
and its joint's range of motion closely matches human capability. Inspired by passive dynamics
walking robots, our design ensures static equilibrium when standing on flat ground.
The center of gravity aligns symmetrically at the hips, providing stability with straight knees.
This is an essential feature for utilizing passive dynamics in active control, leading to more
efficient gait patterns. Both the knee and ankle motors are positioned above their respective
joints and driven by parallel linkage. This design elevates the leg's center of mass,
reduces the lower leg inertia, and minimizes the torque required by the motors. To evaluate the
robot's locomotion capabilities, we developed a deep reinforcement learning policy for walking.
The policy was trained entirely in simulations. Here we show the progress of the training as the
robot learns to walk. We then deployed the trained policy, Zero Shot, on the physical robot.
Here is a demonstration of a sim-to-real transfer at a command velocity of 0.3 meters per second.
This serves as our baseline for locomotion optimization. With the goal of attaining
better energy efficiency in locomotion, we developed a control policy within the RL framework
to leverage passive dynamics. Our passive control policy modulates the joint torques,
and as a result, the joint switches periodically between a passive and active mode.
This allows the robot to utilize its natural dynamics during leg swing, thus reducing the
need for active motor control. To demonstrate the energy efficiency of our passive policy,
we compared the cost of transport between the active and the passive policy. Real-world
experiments show a 31% reduction in the cost of transport at a command velocity of 0.3 meters per
second. The Duke Humanoid serves as an open-source research platform for studying humanoid locomotion.
The integration of passive action into the reinforcement learning control showcases the
potential for achieving more natural and efficient humanoid movements, offering a promising direction
for future research. Please check our website for the open source of both software and hardware.