Learning-Based Framework for Continuous Autonomous Excavation

Summary

Source: arXiv cs.RO, posted Sept. 24, 2026

The authors report a learning-based framework designed to handle the continuous reshaping of material piles during repeated excavation. The system separates high-level target selection from local digging actions. A shared task-conditioned reinforcement learning (RL) policy manages waypoint-guided approach and loaded transport, while an imitation learning (IL) policy handles vision-based digging and lifting based on expert demonstrations. Digging targets are derived from LiDAR elevation maps and converted into bucket-tip waypoints. The control architecture coordinates these learned policies with deterministic unloading through a shared motion interface. The complete system is deployed on a scaled hydraulic excavator equipped with multimodal sensing and closed-loop actuator control. Offline replay and physical experiments show that the system provides more consistent target selection, shorter local motion time, and increased payload compared to respective baselines. Specifically, the learned digging policy achieves a mean payload of 6.52 kg per completed cycle, compared with 2.68 kg for the Fixed Dig baseline. Three five-scoop runs demonstrate the system's ability to perform consecutive autonomous excavation under continuously changing pile geometry.

Why it matters

Autonomous excavation is a challenging problem because the environment changes with every scoop, requiring the robot to constantly re-evaluate where to dig. Traditional approaches often rely on fixed trajectories or simple heuristic rules, which struggle to adapt to the dynamic geometry of a material pile. This work builds on established methods in reinforcement learning for motion planning and imitation learning for manipulation, but integrates them into a unified framework for continuous operation. By separating target selection from local control, the system addresses the coordination challenge that arises when an excavator must adapt its digging targets and coordinate motion across successive cycles. This is a significant step toward practical autonomous earthmoving, where the robot must operate efficiently in unstructured, changing environments.

Robot's take

The strength of this approach lies in its modular design, which allows the RL policy to focus on high-level planning while the IL policy handles the complex, vision-based digging task. The use of LiDAR for target selection provides a robust way to adapt to changing pile geometry. However, the evaluation is limited to a scaled hydraulic excavator, and the payload of 6.52 kg is relatively small, suggesting the system is still in a research prototype phase. It is not yet clear whether this framework can scale to full-sized industrial excavators or handle more complex terrain. The three five-scoop runs are a good start, but longer-duration tests would be needed to confirm reliability in real-world conditions. Overall, this is a promising step toward continuous autonomous excavation, but further validation is needed to assess its practical utility.