Flyped: A Flywheel Actuated Biped for Robust and Efficient Robot Locomotion over Varying Terrain
Here I present a novel wheel-leg hybrid robot morphology that uses a wheel as both an inertial actuator during walking and a drive actuator during rolling. This morphology allows the robot, called the Flyped, to demonstrate more versatile and robust behaviors during legged locomotion when compared to a normal bipedal walking system, while also giving the robot the ability to efficiently roll when it encounters flat, structured terrain.
Generating diverse, dynamically feasible, and efficient locomotion over varying terrain is a difficult challenge, and I tackle this challenge by decomposing the problem into hierarchical sub-planners. At the top level, an implicit graph A* search is performed to decide gait behaviors (rolling vs legged), and initial footstep locations. Then, for each gait mode, a local trajectory optimization problem is solved to produce dynamically feasible center of mass and joint motions of the robot within that mode. Finally, a time-varying linear quadratic regulator is used to obtain feedback joint torques needed to track these state trajectories during execution, which are then passed to high-speed low-level motor controllers for real-time control.
At the highest level of the motion planning framework, an A* search is performed on an implicit graph to generate the location of transitions between legged and rolling gaits, as well as define initial footstep locations for the walking gait. An implicit graph is used due to excessively large memory requirements of an explicit graph in this planning domain. A guided A* search is chosen to guarantee search optimality, as well as speed-up search time through a well-chosen heuristic function. The optimal path is defined as the most energy efficient feasible path from the current robot state to a desired robot world location (locomotion). This kinematic planner is given an initial robot state, a goal robot state, and a height map defining the height of the terrain at every discrete x position in the map and returns a set of actions defining the optimal gait transition and footstep locations throughout the trajectory.
I implemented the graph search planner in c++ for fast planning times, and are able to generate plans for 10 meter long maps discretized into 5 centimeter segments in less than 0.1 seconds.
The high-level gait planner produces footstep plans for the Flyped robot for a given terrain map, but these initial footstep plans are not checked to be dynamically consistent, and thus are only ensured to be kinematically feasible. To generate physically feasible legged locomotion state trajectories over rugged terrain from these initial kinematic footstep plans, we utilize a legged-locomotion specific trajectory optimization parameterization proposed by Winkler. This nonlinear program framework for motion planning of legged systems is called Trajectory Optimization through Phase-Based End-Effector Parameterization, and is effective in simultaneously generating footstep sequences, step timings, foothold locations, swing-leg motions, full body motion, and required contact forces over non-flat, unstructured terrain. This trajectory optimization technique is capable of generating motions for multiple footsteps in only a few seconds over varying terrain. Because the optimization can reason directly about foot contact forces during locomotion, it can generate emergent behavior like full-flight phases, which are essential for highly dynamic movement.
Full Motion Planner
The motion planner demonstrates the ability of the legged locomotion trajectory optimization framework to generate dynamically feasible motion plans over varied terrains. Here a sample motion plan from the full motion planner shows the planner leveraging the efficiency of rolling locomotion to roll through benign sections of the terrain, and then traverse difficult terrain with legged locomotion when necessary. This map includes a variety of typically difficult-to-traverse environments such as stairs, deep holes, and steep ramps. Despite this, the motion planner is able to quickly generate an energy-optimal, dynamically feasible motion plan for the Flyped robot.
To test the capabilities of a reaction wheel-leg hybrid robot, a prototype Flyped robot was designed and fabricated. The robot consists of a lightweight aluminum base chassis with a large diameter flywheel mounted on its central axis. The wheel is coated with a thick high-friction rubber sheath to allow it to roll on the ground with minimal slippage. The wheel is actuated by a high-torque, high-speed brush-less DC motor. A BLDC motor was chosen that satisfied the torque requirement found in simulation experiments of legged locomotion, while also maximizing the flywheel saturation velocity to allow for maximum stabilizing torques during walking. Two lightweight aluminum legs are mounted to the base chassis to enable legged locomotion and are each actuated at the hip and knee by two X5-4 Hebi modules. The ease of use of these modules made them a great choice for quick prototyping of the Flyped system.
The Flyped control software architecture consists of four sub-components integrated together in a ROS package. This ROS package enables seamless communication between the Hebi modules, the ODrive motor driver, an Arduino Teensy, and a control computer.
Initial Hardware Results
This research project is fully supported by the Biorobotics Lab at Carnegie Mellon University. Special thanks to my advisors Matt Travers and Howie Choset.