Neural Network Controller for Autonomous Pile Loading Revised

Wenyan Yang, Nataliya Strokina, Nikolay Serbenyuk, Joni Pajarinen, Reza Ghabcheloo, Juho Vihonen, Mohammad M. Aref, Joni-Kristian Kämäräinen

Introduction

Autonomous pile loading—the process of a machine driving into a pile of material and filling its bucket—remains a "grand challenge" in earth-moving automation. The difficulty stems from the unpredictable physics of material interaction and the extreme variability of outdoor environments. Historically, simple Neural Networks (NNets) were considered too brittle for this task, often failing when conditions deviated from training data. This led many researchers to favor Random Forest (RF) controllers. This study addresses the "generalization gap" by investigating whether a controller trained on "easy" summer data (dry gravel) can survive a "hard" winter environment characterized by ice, snow, and frozen mud.

Proposed Architectures

The researchers developed a revised approach called NNetV2 and two attention-based variants: ANNet (Single Attention) and DANNet (Dual Attention).

  • Scaling Up (NNetV2): The authors replaced the tiny, 5-neuron "shallow" networks found in previous literature with a more expressive architecture featuring two layers of 200 neurons.
  • Neural Attention (ANNet): This module generates a "mask" for input sensors, allowing the network to dynamically prioritize specific signals (like hydraulic pressure) depending on the stage of the loading task.
  • Dual Attention (DANNet): This novel architecture applies attention to both the inputs (sensors) and the outputs (control actions). By masking outputs, the network mimics human operators who rarely perform multiple complex actions simultaneously, focusing instead on the single most important action at any given millisecond.
Comparison of the NNetV2, ANNet, and DANNet architectures.

Datasets and Environment

The study utilized an Avant 635 wheel loader to collect data across two primary regimes:

  • Summer Demonstrations: Two sets were created—$D_{II}$ (52 "ideal" demonstrations) and $D_I$ (72 demonstrations including "non-ideal" runs with half-full buckets).
  • Winter Test Set: Experiments were conducted over 30 days in varying winter conditions, including wet snow and icy surfaces, to evaluate real-world robustness.
The Avant 635 wheel loader performing pile loading in various summer and winter conditions.

Key Insights

The "Hidden" Sensor

A critical discovery was the importance of telescopic joint pressure ($p_t$). While drive transmission pressure is sufficient to detect a pile in summer, wheel slip on ice makes that signal unreliable in winter. The telescope pressure remains stable, providing a robust "impact" signal across all seasons.

Data Quality vs. Quantity

Contrary to robotics conventions that favor "perfect" data, using the full, high-frequency dataset ($D_I$)—which included less-than-perfect human demonstrations—significantly improved the neural network's ability to generalize to unseen conditions.

Interpretability through Attention

Attention masks provided a window into the model's decision-making. The system learned to ignore most sensors while approaching the pile but shifted focus to all sensors the moment the bucket made contact with the material.

Results and Conclusion

The results demonstrated a massive leap in performance. While the original shallow NNet and NNetV2 failed completely (0% success) in winter when using limited sensors and cleaned data, the DANNet achieved a 100% success rate when trained on the full summer dataset. This proves that with proper scaling and attention mechanisms, neural networks can outperform traditional models like Random Forests in highly variable environments. These robust controllers now provide a foundation for future work in Reinforcement Learning, where machines might eventually surpass human performance through self-exploration.

Summary updated

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 heavy-duty machine in drastically varying outdoor conditions and delivering high success rate in winter, being trained in summer.
BibTeX
@inproceedings{yang2021neural,
  author = {Wenyan Yang and Nataliya Strokina and Nikolay Serbenyuk and Joni Pajarinen and Reza Ghabcheloo and Juho Vihonen and Mohammad M. Aref and Joni-Kristian Kämäräinen},
  title = {{Neural Network Controller for Autonomous Pile Loading Revised}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2021},
  url = {https://openreview.net/forum?id=HafZdgnVKCY},
}
Figure