An Investigation of Bipedal Walking
Peter Sand, William Kunz, Oren Laskin

Introduction

Humanoid bipedal walking has the prospect of being a very flexible framework for creating moving machines. Most robots use wheels, treads, or large sets of legs to traverse their environments. These methods of locomotion are often very well suited to particular terrains. Human-like walking does not have the specialized performance of these other forms of motion, but it is very general. A good biped provides a compromise between the complex terrain handling of a four or six-legged robot and the high speed of a wheeled vehicle on flat ground.

The primary challenge in achieving this generality in bipedal walking is dynamics. We call an object statically stable if it can stop without falling over (like a car), while we call an object dynamically stable if it cannot stop without falling over (like a bicycle). Statically stable bipedal walking is relatively easy: a cheap wind-up toy can walk forward on two legs. To move quickly and to deal with complex terrain, it becomes important to work with dynamic stability. In this case, the robot must drive its motion such that it is continually falling towards its next step, but never actually loses control.

Dynamically stable walking has been a challenging robotics problem for many years. In the past, bipedal research has focused on dynamically stable machines that are only capable of falling in certain directions due to external supports. Recently, some groups have started to have success with free-standing dynamic walking. The most well known example is Honda's biped, a robot that took hundreds of scientists working with a very large budget.

The Carnegie Mellon Robotics Club, despite its limited experience and resources, has been experimenting with bipedal walking for several years. As part of a SURG project last year, our biped was able to perform free-standing dynamic walking on a flat surface. The mobility and robustness of this biped was very limited, but it did give us significant experience with the issues involved in bipedal walking.

This year, based on our previous experience, we have decided to take a different approach. A prototype biped has been constructed that is much simpler and lighter than our previous biped. The old biped had many actuated joints; using it we were able to determine which joints were essential and which were not. By eliminating several superfluous degrees of freedom, the mechanical structure of the biped becomes much simpler and therefore much lighter. The simplicity makes physical analysis much more tractable, and the lightness significantly improves the speed and strength of actuation.

Proposal

The current prototype biped consists of five model-airplane servos and a several structural elements. One servo is attached to a vertical beam that can function as an actuated torso. To work with different walking algorithms, this beam can be augmented with weights or completely removed. Each hip has two servos: one for side-to-side actuation and one for forward actuation. By synchronized movement of both side-to-side servos, the feet of the biped can be raised and lowered. The ankles are currently unactuated. In dealing with uneven terrain, we will probably need to add another servo or a passive ankle spring.

We will start by augmenting our current prototype biped with a sensor system. In particular we will use two cameras, one facing forward and one facing to the side. The previous biped used an accelerometer-based tilt meter. This tilt meter was susceptible to deviations caused by the acceleration that necessarily occurs when the biped moves. A vision based system will completely eliminate that error because it always (with a certain latency) gives the actual orientation of the robot with respect to the world. We have already worked out the details of an efficient algorithm to determine orientation using the assumption that human environments generally have vertical lines. We will use two cheap CCD cameras and a PC-104-Plus frame grabber. For now the computing will be performed off-board, but the size of the frame grabber allows on-board computing in the future.

Using the completed prototype, we then intend to investigate walking algorithms. First we will examine the oscillatory side-to-side motion that was the primary focus of last year's investigations. Once we have adequate control over the height of the feet (something which we were unable to attain last year due to the strength-to-weight ratios of the old biped), we should be able to move the biped forward. Once we have reached this point, we hope to concentrate on improved robustness, especially in response to obstacles.

Applicants

Peter Sand is a senior in Computer Science who has perviously worked on the Robotics Club biped projects. Oren Laskin is a senior in Electrical and Computer Engineering. William Kunz is a sophomore in Computer Science with a strong physics background.

Budget

Since the biped is structurally simpler than the previous biped, many mechanical costs have been reduced. However, the vision system is a new and rather non-trivial cost.

Torso Servo $20
Two Forward Hip Servos $50
Two Side Hip Servos $80
Structural Elements $5
Sensoray Model 311 Frame Grabber $365
PC-104-Plus-to-PCI Adapter $210
Two CCD Cameras $150
Total $880