---

---

Goal

Develop and simulate a simple method of locomotion that would allow a robot to climb and descend stairs safely and autonomously.

Prior State of the Art

There have been successful stair climbing robots, but all of them have severe limitations.

Large tracked robots can power their way up stairs, but their movement is not controlled at all - a human would not feel comfortable climbing a flight of stairs next to them.

Humanoid style robots have also successfully climbed stairs Their major problem includes the difficulty of dynamic stabilization Additionally, their control software has not been powerful enough to tackle any staircase autonomously.

Robots with passive mechanical assistance, like EPFL’s Shrimp robot, have also climbed some stairs These robots enter into a phase-lock situation when they attempt terrain or stairs that vary in dimensions from the designed case.

Solution

The final design has two tri-stars on a center body tube The four motors in the system drive the two tails and the two sides of small wheels on the spokes The independent drive configuration allows the robot to separate climb and drive motions, even when on a staircase It can position itself on the current stair to have a perfect approach to the next one When climbing, both tails are used to rotate the body of the robot up the stairs When descending stairs, one tail flips forward and minimizes the amount of uncontrolled falling onto each step.

Challenges

Since this was a purely theoretical problem, there was great latitude in finding a solution Limiting the scope of considered concepts was key.

I faced many technical challenges in creating a dynamics simulation of my design Examples include: writing code that logged all of the required data, quickly storing OpenGL rendered frames, and adding an orthoscopic render mode to Open Dynamics Engine Additionally, implementing the high-level stair climbing controller inside the simulator was non-trivial; though this was made easier because in the perfect world of the simulation distances were known, sensor inaccuracies were not simulated.