In an interview with Sander Olson, Cornell engineer and researcher Hod Lipson discusses adaptive, evolving robots. Lipson’s robots compete and learn, and acquire new skills in movement. Lipson uses simulators to co-evolve both simulators and robots. Lipson’s has already demonstrated a robot that can compensate for losing a limb by modifying its movement.
Question: You have recently created a “robotic scientist”. What do you mean by that?
Answer: The robotic scientist is a program that can discern the mathematical laws behind the data and derive an hypothesis. This machine isn’t really a robot, it is more of an algorithm, and its effectiveness is determined by the quality of the data and the amount of computing power used. It is essentially a data mining tool for scientists, and it should accelerate scientific discovery.
Question: How does the automated scientist work?
Answer: The machine tries to find invariants in the data, and those invariants are key to finding the underlying physics. we are confident that it can substantially speed up the pace of scientific research in many fields.
Question: Your robots learn by evolutionary techniques. How do they accomplish this?
Answer: The vast majority of robots today operate by programming rather than by learnt behavior. Self-modeling robotics involves having a robot internally create models of its operation based on its experience, and compare the efficacy of those models. We have already demonstrated a legged robot that can compensate for losing a limb by altering its movements.
Question: How quickly can your evolvable robots adapt their behavior?
Answer: The robots we demonstrated took a few days to generate their self-model. In general, it depends largely on how fast their CPUs are and how complex the experiences to be modeled are. Given the combination of faster machines and increasingly efficient algorithms, robots should be able to able to adapt to their surroundings and circumstances even faster in the future.
Question: Your research lab at Cornell has done has created printable ornithopters. What can these insectoid robots do?
Answer: We have created usable ornithopters using 3-d printing techniques. We recently published a paper that describes these printed ornithopters. The main advantage of crafting these devices from 3-d printing techniques is that we can quickly construct and analyze a wide variety of performance parameters in order to determine optimal design. These ornithopters currently can stay up for 80 seconds. These ornithopters are powered by lithium-polymer batteries, and we hope to have them fly for several minutes within a few years.
Question: You have also discussed “evolving simulators”. What do you mean by that?
Answer: Much of the exploratory evolution of our machines occurs in a simulator, sometimes through hundreds of iterations. But the simulators themselves aren’t completely accurate, so we need the simulators themselves to improve. As these simulators evolve they become increasingly accurate and specific, able to predict machine behavior with increasing accuracy.
Question: Do your robot learning algorithms reach a point of diminishing returns?
Answer: The learning rate for the machines asymptotically slows down after a while. But there is much that can be done to improve the algorithms, and that is what our lab is focusing on with our research.
Question: Is Cornell engaging in any pure AI research?
Answer: Yes, Cornell has active AI programs underway. The AI researchers at Cornell are working on both robotics and non-robotics applications.
Question: What is the first commercial application you see for your robots?
Answer: Within the next several years, we could see commercial applications emerge that use our learning algorithms – mostly in the area of fault tolerance.
Question: Your are an expert in the nascent field of 3-d printing. How does this work?
Answer: The technology for 3-d printing has actually been around for a while. The scheme involves using an inkjet printer to deposit layer after layer of a material to create 3-d objects. This technique can be used to construct a wide variety of structural shapes. Non-structural components, such as batteries, wires, and motors, are considerably more difficult. That is where we are focusing our research.
Question: How much progress do you anticipate in the field of robotics in the next decade?
Answer: The fields of robotics and 3-d printing are both embryonic, but are experiencing exponential growth. This progress should continue for the next decade, leading both to a plethora of consumer products and to scientific advances.
Question: What about 20 years?
Answer: I see my research in terms of both the body – 3-d printing – and the brain – evolvable robots. In the next 20 years, we will see 3-d printing move from being a niche technology to the main method of manufacture for many products. 3-d printing is inherently more versatile and could be more cost-effective than traditional fabrication methods for “long tail” products. 20 years from now, robots will be using machine learning techniques to model the world and to learn on their own.