Prototype Resilient, Self-Modeling Robots

The main contribution described in the Report "Resilient machines through continuous self-modeling" (J. Bongard et al., 17 Nov. 2006, p. 1118, and the accompanying Perspective, "What do robots dream of?", C. Adami, p. 1093) is a robot that can autonomously recover from certain types of unexpected damage, through an adaptive selfmodel derived from actuation-sensation relationships, used to generate forward locomotion.

Neither the Report nor the Perspective mention that the first resilient, self-modeling machines of this type were built by Alexander Gloye-Foerster (now at IDSIA) et al., who won the 2004 RoboCup in the very fast, small-size league (where human adversaries with a joystick have no chance). Gloye-Foerster et al. equipped the RoboCup robots with self-models based on artificial neural networks, to model current properties of their four-wheel omni-directional drives. They showed that when a robot gets damaged and is no longer able to execute a precise driving pattern, it can heal itself by quickly adapting its model of the relation between motor commands and sensory inputs, and using the modified model to plan and optimize future driving trajectories (1, 2).

References
1. A. Gloye, F. Wiesel, O. Tenchio, M. Simon, IT Information Technol. 47 (no. 5), 250 (2005).
2. See also Resilient, Self-Modeling Robots by J. Schmidhuber
(2006)

The correspondence to the left appeared in Science 4 May 2007: Vol. 316. no. 5825, p. 688, DOI: 10.1126/ science.316.5825.688c

Jürgen Schmidhuber
IDSIA, Galleria 2, 6928 Manno-Lugano, Switzerland & Robotics and Embedded Systems, Tech. Univ. München, Computer Science, Boltzmannstr. 3, 85748 Garching, Germany


Related links:

1. The authors published a response (subscription required)

2. More papers (1990-2007) on our Curious Machines that invent and conduct experiments, actively determining which data to collect in order to improve their self-model.

3. More on Reinforcement Learning with self-models (papers since 1989).

4. More on Robot Learning

5. Resilient Machines

Resilient machine with Continuous Self-Modeling Artificial Curiosity Learning Robots Reinforcement Learning Statistical Robotics
.