 |
 |
|
|
|
|
Framework
Objective: The objective of this project is to develop a method for integration of robotic actions at different time scales, ranging from deliberative long-term planning, to reactive behaviors, to servo control through a unified integration scheme. The DRS framework reduces the burden that the human designer must hand program robot for each different robotic task. The framework systematically integrates a wide variety of sensing capabilities, including vision, speech, text, range and tactile sensing, and action capabilities, including navigation, object manipulation, sound generation, pan-tilt eye control, and text outputs. Approach: The approach is to divide the information hierarchy based on the time scale of information that is required to generate the action output. The deliberative layer, reactive layer and servo control layer generate actions that require context of minutes, seconds, and milliseconds scales, respectively. Thus, the hierarchy is not constructed only according to concept levels or behavior levels. The method integrates both symbolic and numeric sensory inputs and action outputs. It uses a hierarchical model with levels in space and time and uses a probabilistic model to deal with uncertainty. Different learning types are combined: parametric, supervised, reinforcement, and self-learning. Further, it enables a fundamentally new way of constructing robot software: autonomous skill development. The robot is autonomous at not only at performance stage, but also the learning stage, and these two stages are the same. In other words, the robots perform autonomous on-line incremental learning for new tasks or new skills through real-time interactions with the environment (including humans) using its sensors and effectors. This approach of development offers a way of dealing with otherwise daunting burden of human programming for robots, especially for environments that are not fully predictable (e.g., hostile environments) and for cognitive capabilities that are extremely difficult to program by hand, such as vision. The servo control layer uses the perceptive frame based control, so that when obstacles block the action, the controller can resume control without re-planning when the obstacles are removed. Demo clips and presentation:
Funding agency: DARPA |
|
