Learn, hack!

URL

http://events.ccc.de/congress/2004/fahrplan/files/313-humanoid-robots-slides.pdf

File name

313-humanoid-robots-slides.pdf

File size

2.5 MB

MD5

f1ae8a9254a4f6447e7c24a6709cd930

SHA1

eaaa4864b96ed3cc83b50573847ae00f2ea0d6fa

Humanoid robots are investigated, because the human body facilitates acting in a world designed for humans, simplifies the interaction with humans, and allows for intuitive multimodal communication. They are also a tool to understand human intelligence. In the talk I will review some humanoid robot projects and introduce you to our robots Alpha, Toni, Kondo, and RoboSapien. Humanoid robots recently received a lot of attention from the media and researchers. Such robots have a human-like body plan, their sensors can perceive the same modalities as humans do, and their actions resemble human ones. In order to act successfully in a complex environment, these robots should adapt their behavior to the encountered situation, learn, and develop. In my talk, I will first motivate why humanoid robots are worth being investigated. On the one hand, motivation comes from the practical side. A human-like body is the ideal shape for a robot acting in an environment designed for humans. Stairs, doorknobs, light switches, tools, etc. are adapted to human proportions. Signs, labels, and acoustic signals target the human visual and auditory system, respectively. The advantages of humanoid robots are most obvious in the direct interaction with humans. Humans practice multimodal communication via speech, gaze direction, facial expressions, gestures, and body language from early childhood on. This interface is hence ideally suited for the communication between man and machine. Another advantage is that the human-like motion of humanoid robots is easy to predict for humans. This facilitates human-robot cooperation based on non-verbal communication. On the other hand, motivation comes from artificial intelligence (AI) research. Building intelligent artifacts is one important method for AI researchers in their quest to understand intelligence. Many researchers now believe that intelligence cannot exist without a body (Embodiment) and that it shows itself only through interaction with an environment (Situatedness). Building robots that act in complex environments allows for the grounding of symbols, which would be meaningless without this link. Robots are also used to test models of biological systems, e.g. for the navigation of desert ants. If one wants to investigate human intelligence in this way, one has to build humanoid robots. In the talk, I will introduce some humanoid robot projects. In particular, I will discuss walking robots, like Asimo und Johnnie, manipulation robots, like Wendy und Armar, and communication robots, like Kismet und Mexi. I will also report about some activities within our research project "NimbRo - Leaning Humanoid Robots". I will introduce you to Alpha, a full-scale humanoid, and Toni, a smaller biped, under construction in our lab. I will discuss mechanical design, electronics, communication, perception, behavior control, and simulation of these robots. Furthermore I will describe the commercially available robots RoboSapien and Kondo that we augmented with a Pocket PC and a camera in order to make them autonomous. Because humanoid robots are supposed to carry out a large variety of tasks, their evaluation is not easy. One possible evaluation method is the organization of competitions. In order to give you an impression of the state of the art, I will report about this year's RoboCup Humanoid League competitions, which took place in Lisbon. It will become obvious that many research problems need to be solved, before humanoid robots can be used in the real world. Among them are the development of high-performance actuators, sensors, and power supplies. A big challenge is the realization of energy-efficient dynamic walking that supports the dynamics of the system. There is a need for modular behavior control architectures that make controlling the many degrees of freedom of humanoid robots tractable. Furthermore, the integration of individual components must be done in a way that produces synergy effects, such as in audio-visual speech recognition and in active perception. Last, but not least, learning techniques need to be developed to adapt the behavior of humanoid robots to the situation at hand. These methods must be able to learn from few examples. One way to achieve quick learning is imitation. The humanoid body and human-like motion primitives simplify the mopping of observed actions to the robot. This could allow even naive users to 'program' a humanoid robot by demonstration. My talk will conclude with a speculative outlook to the future of humanoid robots. Using them as personal robots could change the way we live significantly.