No abstract

Fig. 1. Overview of the method. A: The initial values of a robot control policy's hidden layer is set by supplying the word2vec embedding associated with a command such as 'stop' to one neuron in the input layer. The policy is then downloaded on to a robot, and the sensor data generated by its movement is supplied to the remainder of the input layer (dotted arrow), further altering the hidden-and motor layers. After evaluation, the robot's behavior is scored against an objective function paired with the command, such as one that penalizes motion. The same policy is then evaluated four more times (two of which are shown in B and C) against four other commands and objective functions. Policies are trained to maximize mean performance against all five of these functions (D). After training, the best policy is supplied with a sixth, previously-unheard synonym of 'stop', and its behavior is scored against the 'stop' objective function (E) (youtu.be/MYegNCJ5bWU).Abstract-Enabling machines to respond appropriately to natural language commands could greatly expand the number of people to whom they could be of service. Recently, advances in neural network-trained word embeddings have empowered non-embodied text-processing algorithms, and suggest they could be of similar utility for embodied machines. Here we introduce a method that does so by training robots to act similarly to semantically-similar word2vec encoded commands. We show that this enables them to act appropriately, after training, to previously-unheard commands. Finally, we show that inducing such an alignment between motoric and linguistic similarities can be facilitated or hindered by the mechanical structure of the robot. This points to future, large scale methods that find and exploit relationships between action, language, and robot structure.

In evolutionary robotics, populations of robots are typically trained in simulation before one or more of them are instantiated as physical robots. However, in order to evolve robust behavior, each robot must be evaluated in multiple environments. If an environment is characterized by f free parameters, each of which can take one of n p features, each robot must be evaluated in all n f p environments to ensure robustness. Here, we show that if the robots are constrained to have modular morphologies and controllers, they only need to be evaluated in n p environments to reach the same level of robustness. This becomes possible because the robots evolve such that each module of the morphology allows the controller to independently recognize a familiar percept in the environment, and each percept corresponds to one of the environmental free parameters. When exposed to a new environment, the robot perceives it as a novel combination of familiar percepts which it can solve without requiring further training. A non-modular morphology and controller however perceives the same environment as a completely novel environment, requiring further training. This acceleration in evolvability -the rate of the evolution of adaptive and robust behavior -suggests that evolutionary robotics may become a scalable approach for automatically creating complex autonomous machines, if the evolution of neural and morphological modularity is taken into account.

Random recombination in evolutionary algorithms can be counterproductive in systems that evolve increasing modularity, because such operators do not preserve community structures during their development. Partly because of this, methods have been proposed that derandomize recombination by placing potential crossover locations under evolutionary control. Since crossover is likely to be particularly useful when genetic material that generates incipient phenotype modules is recombined, there may be an advantage to seeking such modularity directly in the phenotype and probabilistically focusing recombination at such "hotspot" locations. Here we show that such phenotypically-aware crossover operators can outcompete random or evolved crossover points as the size of the system being evolved grows. As this crossover operator can be viewed as epigenetic, and as epigenetic processes seem to be common in biological systems, other such epigenetic mechanisms may further improve future evolutionary algorithms.

A long time goal of evolutionary roboticists is to create everincreasing lifelike robots which reflect the important aspects of biology in their behavior and form. One way to create such creatures is to use evolutionary algorithms and genotype to phenotype maps which act as proxies for biological development. One such algorithm is HyperNEAT whose use of a substrate which can be viewed as an abstraction of spatial development used by Hox genes. Previous work has looked into answering what effect changing the embedding has on HyperNEAT's efficiency, however no work has been done on the effect of representing different aspects of the agents morphology within the embeddings. We introduce the term embodied embeddings to capture the idea of using information from the morphology to dictate the locations of neurons in the substrate. We further compare three embodied embeddings, one which uses the physical structure of the robot and two which use abstract information about the robot's morphology, on an embodied version of the retina task which can be made modular, hierarchical, or a combination of both.