Using probabilistic reasoning over time to self-recognize

Using the probabilistic methods outlined in this paper, a robot can learn to recognize its own motor-controlled body parts, or their mirror reflections, without prior knowledge of their appearance. For each item in its visual field, the robot calculates the likelihoods of each of three dynamic Bayesian models, corresponding to the categories of “self”, “animate other”, or “inanimate”. Each model fully incorporates the object’s entire motion history and the robot’s whole motor history in constant update time, via the forward algorithm. The parameters for each model are learned in an unsupervised fashion as the robot experiments with its arm over a period of four minutes. The robot demonstrated robust recognition of its mirror image, while classifying the nearby experimenter as “animate other”, across 20 experiments. Adversarial experiments, in which a subject mirrored the robot’s motion showed that as long as the robot had seen the subject move for as little as 5 s before mirroring, the evidence was “remembered” across a full minute of mimicry.