The agent deploys novel strategies to clear the deformable tree branches around the power line using the whole arm.

The agent uses novel strategies to clear deformable branches in an agricultural setup, allowing an external inspection camera to view a partially or fully occluded target (e.g., fruit, diseased regions, or a navigation lookout).

Strategies exhibited by our approach in the real world operating on multiple tree branches of varying species.

Another real-world strategy is shown alongside the segmented point clouds and the real-time change in the occlusion heuristic and contact probabilities.

We borrowed parametric L-system rules, dynamics parameters, & actuation details from [1][2] to generate realistic, self-similar branching structures.

We run thousands of parallel tree models in the NVIDIA Isaac Gym physics simulator, incorporating diverse geometric structures, dynamic behaviors, and contact patterns via domain randomization.

• We present a multi-modal, whole-arm contact policy learning framework to manipulate deformable clusters, combining both point clouds (from an RGB-D camera) and touch detection inputs (from proprioceptive sensors).  

• We present and experimentally validate an efficient distributional state representation of the complex deformable cluster geometry that is efficient for both RL training and inference.

• We leverage a novel context-agnostic occlusion heuristic and reward strategy to clear deformables from a target region. Our implementation and primary focus is to clear power lines (validated in both simulation and real) of overhanging tree branches, but generalisable to other applications .

• We demonstrate a zero-shot sim-to-real policy transfer that handles unseen branch topology and displays novel clearance patterns in real.