ATLAS take home the Gold!
At Robogames 2009, ATLAS represented DASL in the autonomous humanoid dash. After 2 days of tuning and trials, ATLAS completed the course in slightly over 1 minute, winning 1st place (by default). Here is a fun overview video of the project, so watch and enjoy.
Walking Algorithm Progress
Here is a short video of our IK solver producing a simple walking trajectory in MATLAB. The solver was made using AutoLev’s dynamic software, and will be run in realtime on the robot to convert high-level commands into actual joint angles. The purpose of an inverse kinematic solver is to convert a desired trajectory of points on the robot (such as the feet) to a series of joint angle commands that the motors can use. Using a basic walking trajectory developed by the Hubo lab (Park et. al), the resulting trajectory plot for a sconstant speed walk looks like this:
Making an effective open loop trajectory required significant trial and tuning. Despite the obstacles, however, the IK solver & walking pattern was a success. The first steps of the ATLAS robot are documented in this video.
During our experiments, howerver, we discovered critical complicating factors we found included:
- Joint compliance at each servo allowed position errors at the feet of up to 15 mm from the desired lcoation. The difference between the loaded leg and the swinging leg often meant that the feet landed crooked, interrupting the gait.
- Transmit-to Receive switching in the USB to dynamixel adapter was very slow, leading to a total motor communication speed of less than 500Hz. With 15 motors to command, the effective update speed was less than 30Hz, cause noticeable vibration as the servos repeatedly accelerated and decelerated.
- The grip of the feet was unpredictable on rubber and linoleum, requiring high-friction “shoes” to be added.
- The relatively wide hips cause the body to “sag” slightly to the inside during single support, requiring compensation in the hip trajectory to stabilize the walk.
Project Risks & Mitigation
Accurate sensing of Zero Moment Point
Keeping the ZMP within the support area is critical to maintain stability of a biped robot on a flat surface. Since the zero moment point corresponds to the center of pressure of one or both of the feet, the applied forces and moments at each ankle must be measured. A brief trade study of force sensors showed that simple capacitive force transducers at each corner of the foot can effectively sense COP location.
Measurement of angular momentum and body orientation
During biped locomotion, one of the most critical phases is foot landing. The swinging leg/foot must land to counter the forward fall of the body. If the step is too large, the robot will lose too much forward velocity, stopping suddenly and likely falling over. Conversely, too short a step will cause the robot to trip and fall. To land within this “sweet spot”, the angular momentum can be calculated from the pitch rate, and used to control foot landing. A similar circumstance is true for sidestepping.
Without a way to measure this, the robot’s gait will likely be slow and shaky. An IMU will be used to provide the necessary feedback to make the walking gait more robust
Robot must recover from a fall and stand up on its own
To stand up after falling over, the robot must be able to roll onto its feet, balancing even as it lifts the body from the floor. Careful analysis in CAD will be performed to ensure that the center of gravity will remain over the support parallelogram during this maneuver. This restriction will be used to constrain the mechanical design.
PicoITX Mainboard lacks hardware interfaces for sensors
Proper integration of all of the sensors presents a large risk because multiple analog and serial ports are not present on the typical ITX motherboard. This necessitates additional hardware and/or software to work around. Rather than design a serial line bus, or depend on DAQ hardware, the ethernet protocol will be used to link the main board to perpheral microcontrollers. Example C code exists for analog to digital conversion and simple ethernet communication for many microcontrollers, eliminating much of the development time. The robustness and expandability of the ethernet protocol will make adding or changing sensors easy. See the Computer Page for more details.
Scanning the Competition
A bit of investigation shows that many of the robots of previous years far exceed the capabilities of the Robonova or Bioloid.
In 2008, the stair climbing champions climbed up and down a short flight of stairs:
The winner of the 2008 Kung Fu competition is linked below. One surprising thing about Kung fu is how much fighting uses scripted motions. All of the gestures seem to be pre-programmed. This suggests that a good control system and movement scheme could produce more exatc motions, better-aimed punches, and quicker recoveries.
2007 Kung Fun Champions
Robo-One Grand Champion at IREX
These videos show the variety of hardware out there just as conspicous is the need for better control methods. As an example of what good control can do, here is a demo video at the ROBO-One 10 copeition in 2006:
The KHR-2 HV shows a smooth dybanic gait, rapid side-stepping, and excellent balance control. Even though this platform is only “pro-sumer” level, it clearly looks more human-like and controlled than many of the entries above.