- Download 3
- File Size 284.28 KB
- File Count 1
- Create Date November 11, 2021
- Last Updated November 11, 2021
D3.3 Report and Software for using interaction primitives for online learning to add tasks in real-time (Module#8) - ABSTRACT
This document reports the advancements developed in the context of WP3, describing the work conducted and the achievements accomplished during TUDA’s work on task T3.4 “New task configuration through learning from demonstrations”. This report closes the integration for Module #8: Robot Motion Planning based on learning from demonstrations. The aim of WP3 is to provide the robots reactive and efficient decision-making capabilities to adapt to highly unstructured environments, as the ones in which the human operator might constantly enter in and modify the workspace. In a situation in which an operator enters in the workspace, the robot should be able to efficiently recompute its future actions to safely adapt to the human. To do so, in WP3, a set of algorithms have been developed for human-aware task and motion planning. To safely adapt the robot motions to humans, dynamically adaptable, human-aware, motion planning algorithm has been developed, in collaboration with CNR. The algorithm has two main elements. On one hand, a human occupancy prediction algorithm has been developed. Given the current human state, the algorithm predicts the probability density function on where the human will be in the future. Then, the motion planning algorithm, developed by CNR, takes this information into consideration to plan safe motions in the workspace. During task T3.4, the problem associated with modelling and learning the human occupancy prediction modules has been addressed. Based on highly expressive probability density function models, known as Normalizing Flows; in our work, we explored the integration of these models for the problem of learning conditioned human occupancy predicted densities. Additionally, we have explored the possibility of using previous data to guide the search in motion planning and improve its efficiency. By using CNN and ANN as approximators, we trained predictors on the feasibility of some key decisions in robot grasping and further derived heuristics from those predictors to accelerate motion planning. All above contributions are detailed in dedicated sections of this document. Further technical information is available in relevant disseminated project results, attached as papers in the Appendix.