The Artificial Intelligence group (AI Lab) is a research unit inside Softbank Robotics Europe which focuses on developmental robotics. Our goal is to make progress in the understanding and modeling of the mechanisms of development and learning in robots.
In this context, three topics are proposed for mid- to long-term internships. We aim to offer a stimulating environment, working on high-tech humanoid robotic platforms. The following internship proposals are broad so that they can be discussed and oriented depending on the profile of the candidate.
1 - Learning to identify traversable terrain for Pepper (Applied topic)
In recent years, image processing algorithms made impressive progress, in particular through the use of Deep Learning methods. These approaches have been applied in the context of mobile robotics to evaluate what part of an image is traversable and what parts represent potential obstacles. The goal of this internship is to implement such an approach to allow the Pepper robot to identify in which direction it can safely move.
Taking inspiration from already existing works in the literature, the intern will develop and implement a deep network to process the sensory flow provided by the robot’s camera, using Python and TensorFlow. As data is the major fuel of Deep Learning approaches, a part of the internship will also consists in evaluating the best method to collect and provide labelled data to the machine learning algorithm (manually label data, use a stereo/depth sensor, use Pepper’s bumpers...).
2 - Learning to see as a sensorimotor predictive model (Applied and research topic)
The AI Lab develops a new approach of perception grounded in the sensorimotor experience of the robot. This new paradigm differs largely from the perspective encountered in traditional artificial perception and requires the development of new models and algorithms. The lab applied this approach to the problem of vision and developed a sensorimotor model allowing a naive agent to explore its environment and discover how sensory inputs provided by its camera change when it moves. This model can later be used by the agent to predict its potential future sensory inputs, as well as the best action to fulfill a visual task. The goal of this internship is twofold, with both an applied and a fundamental aspects:
1) The model takes inspiration from human vision and the covering of our retina by numerous heterogeneous receptive fields which process only small local visual features. Similarly to what can be observed in the brain, this data needs to be processed in a parallel fashion. The first objective of the intern will be to develop a parallel version of the existing serial algorithm, and to run it on a powerful GPU in order to greatly improve its computational efficiency.
2) The second objective is more fundamental and research-oriented. The predictive model learned by the robot captures the properties of vision. Guided by its supervisor, the intern will analyze more in detail the structure of the predictive model. Depending on the intern’s interest, the model could be used to characterize visual properties from a sensorimotor perspective, or to solve visual tasks which require the robot to interact with its environment.
3 - RNN with gated transitions (Research topic)
Recurrent neural network (RNNs) are a type of neural networks that have feed-forward as well as recurrent connectivity. Recurrent connections are used to transfer neural information from one timestep to another future timestep. RNNs are mainly used today in Deep Learning to learn the temporal structure of data. For example, they are used to learn how to generate sequences of characters, or words. In the field of developmental robotics, and especially following the approach of predictive coding, we would like a robot to learn to predict its future sensor values based on his current sensor values and a sequence of motor commands.
In our experiment, a robot navigates in a closed environment, and learns to predict the future values of its distance sensors based on its motor commands. The goal of this internship is to implement a variation of classical RNNs, where the weights that transfer the neural activation across time are gated using latent variables. We believe that these latent variables can be used to represent properties of the environment that can influence the prediction of the future sensory state.
Engineering Diploma / Master of Science / Master of Research
Language: English is mandatory, French is optional
Programming language: C/C++, Python, or Matlab
Please send your application with the following elements:
- Internship topic(s) you’re interested in,
- Short paragraph about your expectations for this internship,
- Brief summary of your relevant experience,
- Reference of this job offer.
Vous vous reconnaissez dans ce poste, alors transmettez-nous vite votre candidature (CV+LM) en indiquant la référence INN/IALAB/RG.