Experienced excavator operators are familiar with a trick that allows them to throw the bucket's contents beyond the reach of the arm. The neural network-equipped robotic excavator proved to be a diligent learner, capable of the precise throwing of stones further than the reach of the arm. Next up is throwing bulk materials and improving accuracy for working with the bucket at different heights.
Researchers from Switzerland (ETH Zurich) reported on the training process of the robotic excavator's neural network for precise manipulation of bucket contents. A reinforcement learning-based neural network was trained to throw balls and stones to a designated point that was beyond the reach of the arm (up to 9.5 m with a reach of 7.5 m). Such operations will help robotics tackle a wider range of tasks with lower energy costs for movements, as well as make their work safer.
The excavator performed grabs and throws with two degrees of freedom, which was not rigidly fixed to the arm. Throws were made straight with just the arm working, as well as with a cabin rotation. In the second case, accuracy was slightly lower, but in any case, the projectile deviated from the target point by no more than 30-40 cm.
Researchers trained the neural network on a modified 12-wheeled Menzi Muck M545 excavator. Previously, they had trained the excavator to perform a number of non-trivial operations, such as teaching it to build a stable wall from unprepared stone blocks. The excavator assessed the balance of the stones itself and constructed a sturdy stone enclosure. For precise autonomous operations on the terrain, the excavator builds a model of the surrounding space using sensors installed on it, in which it performs the assigned tasks.
Source:
Source: 3dnews.ru
