Dr. Hod Lipson from the Creative Machines Lab Discusses “Robotic Scientists” at TCNJ

On Wednesday, April 20th, The School of Engineering and School of Science sponsored a presentation by Dr. Hod Lipson, Professor of Mechanical Engineering at Columbia University, who spoke to a group of students and faculty at TCNJ about his research on self-aware and self-replicating robots. He is an award-winning researcher and director of the Creative Machines Lab where he discovers new ways to make machines that are creative and innovative.
Lipson focuses his research on evolutionary robotics, a branch of robotics that uses processes inspired from biological evolution to “breed” new robots, rather than design them manually. He started his lecture with a story of how he got to where he is now at Columbia, saying that it took years and “a lot of serendipity and research.” While researching robotics, Lipson learned the one major weakness of all robots- their inability to adapt to change. From then on, he made it his main goal to create robots that are self-aware and are able to adapt to the world around them.
The first two approaches Lipson discussed to building these self-aware robots were the “adapting in simulation approach,” where the robotics scientists evolve the controller in a virtual simulation first then try it in reality, and the “adapting in reality approach,” where the scientists evolve the controller in reality with no virtual simulation. After trying both of these approaches to building new robots, Lipson found problems with each; the first approach resulted in a “simulation-reality gap,” as the virtual simulations did not perform the same in reality, and the second approach took too much time and resulted in worn-out robots.
Instead, Lipson combined the approaches to make the “simulation and reality approach,” which is essentially a cycle of evolving and collecting data. The first step is to evolve the virtual simulator, then evolve the robots, then try it in reality and collect sensor data (what the robot does and feels,) then take that data to breed better simulators, and continue the cycle until all the necessary data is collected.
Lipson used this approach to build a four-legged robot, which had to be self-aware and figure out how to move on its own. It took a few days of trial and error for the robot to learn about itself, as Lipson put it, “It does not know what it looks like, it wouldn’t even know if it was a snake or a tree,” but after four days it figured out that it had four legs. Lipson played a video of the robot forming a self-image, and the audience watched in amazement as the robot learned how to move forward on its own.
Next in the video, Lipson decided to test how the robot performed with damage recovery by removing one its four legs. Remarkably, the robot’s dynamics changed, and it adapted to find a new way to move forward without the leg. In an interview on Columbia.edu, Lipson discusses what it felt like to watch a robot he built be self-aware, “It’s always surprising to see new systems evolve or learn on their own. Seeing a robot learn to do something you didn’t program it to do is a pretty amazing experience.”
Lipson also discussed the artificial intelligence software he created called Eureqa, which finds equations and mathematical relationships in data. The software is available for anyone to download, and is used by thousands to detect difficult calculations in big data. Machines with this software can formulate hypotheses, design experiments, and interpret the results to discover new scientific laws.
Robots and machines that can model themselves have huge practical implications in the real world, says Lipson, and he hopes to continue to evolve self-aware robots over time. He believes that he is on his way to creating the holy grail of robotics, which would be a robot that is self-aware right out of a 3D printer, “I don’t know if we’ll get there in my lifetime,” Lipson says, “but we’re on the path that will eventually lead there.”