Vision-based Automated Loading and Dumping (VALD)

Important results from the project

The goals of the project were to (1) demonstrate autonomous short-cycle loading based on camera vision and machine learning (ML), (2) increase the knowledge and skills of Volvo CE in applying ML models, and (3) establish a collaboration between Luleå University of Technology and Örebro University. Objective (1) has not been fully achieved. However, knowledge of challenges linked to the goal has been identified and development methods based on simulations and transfer to a real environment have been developed and partially validated. Goals (2) and (3) have been achieved in the project.

Expected long term effects

The project has resulted in lessons learned about challenges associated with using cameras to automate navigation in the short cycle loading cycle. In addition, the project has identified difficulties in using machine learning models with reinforced learning linked to positional data for such navigation. The project results suggest that an agent trained in a simulation environment with reinforcement learning can bridge the gap between simulation and reality to complete a simplified version of the navigation task.

Approach and implementation

The project has been carried out in collaboration between Volvo CE (VCE) in Eskilstuna, Luleå University of Technology (LTU), and Örebro University (ÖRU). The collaboration has built on frequent, typically weekly, online meetings between LTU and ÖRU, and monthly reconciliations with VCE. In addition, meetings have been held between VCE and LTU to prepare and carry out experiments. In the latter part of the project, successful methods have been established to carry out experiments in Eskilstuna from Luleå. This has been very important and enabled effective collaboration.