Learning-based framework enables continuous autonomous excavation on a scaled hydraulic excavator, averaging 6.52 kg payload per cycle versus 2.68 kg for Fixed Dig

The framework separates target-conditioned motion from local digging: a shared task-conditioned RL policy controls waypoint-guided approach and loaded transport, while an IL policy learns vision-based digging and lifting from expert demonstrations. Compared with treating excavation as a single control problem, this division lets motion control and digging skill each use a learning paradigm suited to its role. The abstract describes the architecture and the division between the two learning paradigms, and states that the complete system is deployed on a scaled hydraulic excavator with multimodal sensing and closed-loop actuator control.

Digging targets are selected from LiDAR elevation maps and converted into bucket-tip waypoints for motion control, so the digging location is updated as pile geometry is repeatedly reshaped. Target selection is driven by terrain awareness rather than a fixed digging point, directly addressing the problem that repeated excavation continuously reshapes pile geometry. The abstract states that targets come from LiDAR elevation maps and are converted into waypoints, and that offline replay and physical experiments report more consistent target selection.

The control architecture coordinates the learned policies and deterministic unloading through a shared motion interface, allowing successive excavation cycles to be chained. Making unloading deterministic and sharing a motion interface with the learned policies is the engineering integration that extends single digging into a continuous work cycle. The abstract explicitly describes a shared motion interface coordinating the learned policies and deterministic unloading, and states the complete system is deployed.

In physical experiments the learned digging policy achieves a mean payload of 6.52 kg per completed cycle versus 2.68 kg for Fixed Dig, and three five-scoop runs demonstrate consecutive autonomous excavation under continuously changing pile geometry. Relative to the Fixed Dig baseline, it reports higher payload per cycle and validates cycle-to-cycle continuity with consecutive multi-scoop runs. The abstract gives the 6.52 kg versus 2.68 kg mean comparison and the three five-scoop runs, and also reports shorter local motion time.

The result targets autonomous excavation settings where pile geometry keeps changing across repeated cycles, on hydraulic excavation platforms equipped with LiDAR elevation maps, multimodal sensing, and closed-loop actuator control; the layered design (target selection, task-conditioned motion, vision-based digging, deterministic unloading) offers reusable interface ideas for extending similar learning policies to other earthmoving cycles.

The abstract does not state the number of physical experiment repetitions, the spread of the payload statistics, the specific setup of the Fixed Dig baseline, or how the 'more consistent' target selection is measured; the relative contribution of offline replay versus physical experiments is also not separated. In addition, when conclusions from a scaled platform are transferred to full-size excavators, differences in hydraulic response and sensing scale may raise new questions that need confirmation in the main text.