No abstract

No abstract

We have recently proposed two pile loading controllers that learn from human demonstrations: a neural network (NNet) [1] and a random forest (RF) controller [2]. In the field experiments the RF controller obtained clearly better success rates. In this work, the previous findings are drastically revised by experimenting summer time trained controllers in winter conditions. The winter experiments revealed a need for additional sensors, more training data, and a controller that can take advantage of these. Therefore, we propose a revised neural controller (NNetV2) which has a more expressive structure and uses a neural attention mechanism to focus on important parts of the sensor and control signals. Using the same data and sensors to train and test the three controllers, NNetV2 achieves better robustness against drastically changing conditions and superior success rate. To the best of our knowledge, this is the first work testing a learning-based controller for a heavyduty machine in drastically varying outdoor conditions and delivering high success rate in winter, being trained in summer.

One of the key challenges in implementing reinforcement learning methods for real-world robotic applications is the design of a suitable reward function. In field robotics, the absence of abundant datasets, limited training time, and high variation of environmental conditions complicate the task further. In this paper, we review reward learning techniques together with visual representations commonly used in current state-of-the-art works in robotics. We investigate a practical approach proposed in prior work to associate the reward with the stage of the progress in task completion based on visual observation. This approach was demonstrated in controlled laboratory conditions. We study its potential for a real-scale field application, autonomous pile loading, tested outdoors in three seasons: summer, autumn, and winter. In our framework, the cumulative reward combines the predictions about the process stage and the task completion (terminal stage). We use supervised classification methods to train prediction models and investigate the most common state-of-the-art visual representations. We use task-specific contrastive features for terminal stage prediction.