Autonomous Earthmoving: Classical Control Guidance Makes Long-Horizon Trenching Reliably Trainable

Published on .

This paper was presented as a poster at the 5th Workshop on Future of Construction at IROS 2026, in Pittsburgh. All authors are part of my team, the Vortex Studio team at CM Labs.

We look at excavator trenching, a representative earthmoving task with a long horizon: how much soil the bucket takes when digging sets the load for every stage that follows. We split the trenching cycle into six goal-conditioned stages and train in Vortex Studio, simulating multibody dynamics, hydraulic actuation and deformable terrain. At each step, the action blends the command of an untuned resolved-rate inverse kinematics controller with the output of a Soft Actor-Critic policy. The policy's share of authority is either fixed for a whole run or annealed from the controller to the policy during training, an approach we call Action Blending Curriculum over Residual Reinforcement Learning (ABC-RRL).

Neither extreme is reliable on its own: the controller alone never completes a cycle, and unguided reinforcement learning converged on only one seed out of three. With classical guidance, every seed converges, and the curriculum keeps that reliability once the policy has full authority, reaching 94.7–97.3% success.

More details are available on the project page, and the paper and its reviews are on OpenReview.

Poster

Download Open original

Cite

@inproceedings{autonomoustrenching2026,
  title={Autonomous Earthmoving: Classical Control Guidance Makes Long-Horizon Trenching Reliably Trainable},
  author={
    Fonseca Cruz, Evelin and
    Hewlett, Joe and
    Kovacs, Laszlo and
    Gottipati, Sai Krishna and
    Bognat, Adam and
    Hoang, Manhtriet and
    Mars, Clod{\'e}ric
  },
  booktitle={5th Workshop on Future of Construction at IROS 2026},
  year={2026}
}