Des robots pour quoi faire ?

Skilligent is a leader in commercial robot learning and social interaction technology. Their product called Robot Learning and Behavior Control System is a complete robotic intelligence system for learning multi-task service robots with the capability of utilizing robotic arm manipulators.

 

The philosophy of the product is to eliminate programming by introducing a fully trainable control system. After integrating the software into a robot (this still might require programming), a user starts communicating to the robot via gestures, attracting its attention to objects or places, and invoking "reflexive behaviors" through known stimuli. 

 

 

By interacting with the robot and making demonstrations, the user can develop new behaviors and new stimuli in the robot. The iterative training process keeps going until the robot is trained enough to do its job autonomously.

 

 

 

Skilligent developed a robot vision system capable of recognizing objects under real life conditions from a large database of known objects. The same vision system recognizes gestures used for robot training, and provides multiple bottom-up visual attention signals. A bottom-up visual attention system is capable of selecting objects that might be relevant to the current task being learned. The object representations are then stored in an image database integrated with a database of known skills and behaviors.

 

 

A visual robot navigation system that comes with the product allows a robot or a UAV to navigate in known environments. The environments are explored and mapped out during a training session. Also, there is a way to create digital maps from existing images of a territory (e.g. for aerial image-based navigation). The navigation system uses a sequence of observed visual landmarks for position triangulation. Between the observations, the system uses a dead-reckoning mechanism to estimate the robot's position.

A combination of the generalization algorithms is the key to the robot learning technology. For example, during a robot training session, a robot learns from a human how to perform a task. The task learning algorithms build a generalized task representation as a sequence of steps, decision points, conditions, and hierarchical skills. The generalized representation is totally hidden from the end user. The only way for a user to check that the robot's internal task representation is correct is to actually instruct the robot to perform the task autonomously.

 

While observing how the robot acts autonomously, the trainer checks if the generalized task representation is correct ("the robot has learned what it needs to do". If not, the trainer can interrupt the robot at any time and make additional demonstrations. The additional input causes the task learning algorithms to update the internal generalized task representation. The iterative process goes on until the robot "learns what it needs to do;" in technical terms, this means that the generalized task representation causes the robot to behave in a way that matches the expectations of the trainer.

 

This iterative process could be called robot programming by demonstration, but as the user does not have to know what is really going on in the robot's software, the term "programming" might not be fully appropriate. Instead, Skilligent uses the term "robot training," which underlines the fact that the user does not really have to know how the software works in order to successfully train the robot. This quality ensures that a non-professional user (such as a warehouse worker) can successfully train the robot.

The Skilligent software product is designed for multi-task service robots. This means that the robots can be adapted to new tasks without reprogramming. The robot learning technology is the key factor that makes it possible. The software relies on multiple artificial intelligence algorithms working in parallel. It requires a pretty powerful CPU, such as Intel Core Duo, to allow the robot to work on the same timescale as a human trainer. This implies that a robot that uses the software must be controlled by a laptop-class motherboard or a wirelessly connected desktop computer.

The second major product of Skilligent is called Fault-Tolerant Control Framework, a JAUS-compliant "pure" robotic software platform targeted at fault-tolerant vehicle control systems. The platform is said to be originally designed as a software framework for highly redundant avionics and communications systems.

 

 By Michael Somby ( Cette adresse email est protégée contre les robots des spammeurs, vous devez activer Javascript pour la voir. Cet e-mail est protégé contre les robots collecteurs de mails, votre navigateur doit accepter le Javascript pour le voir )