The project primary goal is to develop a set of computer algorithms and create SDK (software development kit) aimed at robot "thoughts" generation and transforming them to language (text), ready to be spoken out by a text-to-speech device. By means of "thoughts" robots will be able "to talk" without hardcoded phrases, express their intrinsic multi-dimensional state in language sentenses resulting in more efficient and more natural cooperation with human. The set of algrithms must be efficient, fast and simple enough to be implemented on the real-world, limited-capabilities robot systems. Mathematical instrument for algorithms synthesis for this SDK are artificial neural networks due to their ability to learn, store knowlege, generalize, perform non-linear mappings in the multi-dimensional spaces. Neural network algorithms can be efficiently implemented in hardware (or emulated by simple software), providing low-cost and fast speed solutions.

Most of research in autonomous robotics now is devoted to robot body control, their ability to walk, jump, recognize visual objects, hear voice, synthesize speech and do various things. There is good progress there.

However, real usefulness of robotics (or general machines) applications for people requires two important things that for now may be considered missing:

(a) ability for robots to think / feel

(b) ability for robots to communicate their thoughts to human as we , human, do by language

To make these things reality a concept of artificial neural networks (NN) can be applied, because NNs are good at non-linear multidimensional space mappings, and, as it is known, our words and language constructs can be described in terms of multidimensional vectors, where certain complex concepts are decomposed onto the basic concepts (the basis in this space). In addition well-known properties of NN help such as

- NNs (unlike deterministic algortitms, such as linguistical AI) deal fine with missing, incomplete and noisy data

- NNs can learn from examples just like humans

- NNs store knowlege in a compressed form which is essential for modelling the world in a finite capabilities device such as autonomous robot

- NNs are good approximators and thus extrapolators, which is essential for forecasts and robot planning activity.

- NNs are naturally perfect for implementing reactive behaviour (like reflex), because they are very fast, once being trained.

Our assumption is that "thought" is in the strong connection with "language" and being able to model concepts and sensefull un-predetermied text constructions generation is a half way to make thinking robots.

The project will result in development of a (C/C++) Software Development Kit containing the set of ready algorithms to deploy in particular robots). "Thoughts SDK" will represent an essential tool for robot interaction with human. Humans must know "what is going on" in the robot and robot must be able to report it's condition or ask favours from human to complete the common goal when cooperation of robot and human is required. Thoughts SDK will make robot-human interaction more natural that will extend robotics applications usage by non-technically oriented people, such as small kids, grantparents, ill people, and so on. In particular, Thoghts SDK will facilitate:

- concepts representation and analysis
- text constructs synthesis for communication with human relative to the robot current state, history, estimated future - aimed at the cooperation between robot at human to satisfy the common goal.
- internal robot "thinking" software architecture (represented in the set of developed software modules)

Secondary goal is to develop a hardware prototype that will use the "Thoughts SDK" and demonstrate the possibility of using these aglrotihms on the autonomius robotic system (compact device with limited capabilities, unlike supercomputers used to run AI algorithms). The prototype will be a compact, low-cost talking robot that does not have a set of pre-defined phrases, rather, the phrases will be generated from logically-connected subtemplates. The robot will be able to talk about it's state, about the environment and host state, make "wishes" and "forecasts". There could be a wide spectrum of possible applications derived from such prototype, from toys to old-people social helpers.

Requirements:

• interest in the subject
• math (linear algebra, calculus)
• neural network theory and practice
• programming skills (C++ / Java - to be determined)
• interest in AI / robotics field (experience or customers are appreciated)
• basic language construction knowlege

Optional

• real hardware robotics experience
• speech synthesis
• speech recognition
• image/ object recognition

PROFILE 1: Elaborate initial math task definition, perform experiments, collect training data, code algorithms, compare and report results

PROFILE 2: Run algorithm tests on their real existing hardware robot implementations

Our research will start from basic multidimensional spaces study where the dimesions are concepts. Several approaches we have elaborated that must be independently verified and the best variants should be chosen. Each partner in the group would study the different varian of the neural network task definition, peform the training sets construction, experiments, report results and make the comparison. The group will share whatever common thing could be used by more than one partner in the group. 

The work breakdown idea is in "equalness" of the Partners in the sense that each partner will study the specific task definition. There can be several ways in which the task can be formalized and each way deserves it's own development, test and comparison of the results.

Additional (special) Partners may join the project in the case of the particular expertise to contribute (like hardware speech synthesis - that would be cool to deploy and test), based on their additional idea that they may get reading this proposal.

We see the following basic steps of the project

(a) coordinating the common goal and the set of task denifitions contributed by the parnters that worth verifying

(b) assign each task definition to each Partner, except for special-parnters that would be assigned related separate tasks such as hardware speech engine tests or speech recognition tests

(c) collectively study all task definitions, compose algorithms, aggregate data, test and compare with other methods generated by all Partners. During this work useful data, algorithms and ideas will be shared between Partners to raise the efficiency of the study

(d) make common conclusions , create final library of algorithms, reports/papers, experimental data, test-cases, documentation and deliver this as an SDK that can be commercialized and/or used for the selected future directions

We're looking for partners with whom we can separate this complex task into subtasks to speed-up investigations, share experience and useful common modules developed and what is most important - assess different approaches and compare the results to select the best direction of the research in this area.

• Research & Development centers in the Universities / Institutes who perform investigations in the applications of neural network to text / speech processing
• private (Corporate) research centers who perform projects in robotics/ neural networks / intellectual text processing-generation fields
• Institutions interested in using the results of the project in their robotics research or commercial applications