George Sideris

The work trains one network that classifies which observed point to grasp and predicts depth and yaw there from unsegmented point clouds, learns to clear 200-log packed piles in simulation via behavior cloning (BC), its reinforcement-learning fine-tuning (BCRL), and RL from scratch, and runs twelve field trials on a trailer-mounted hydraulic forestry crane: BC and BCRL deposit 93.8% and 88.9% of pooled inventory against 80.4% for a geometric heuristic, and BCRL succeeds on 83.6% of cycles versus 79.6% for the heuristic and 65.7% for BC, while fine-tuning's simulated stability gain does not carry over to the crane testbed.

The work trains one network that classifies which observed point to grasp and predicts depth and yaw there from unsegmented point clouds, learns to clear 200-log packed piles in simulation via behavior cloning (BC), its reinforcement-learning fine-tuning (BCRL), and RL from scratch, and runs twelve field trials on a trailer-mounted hydraulic forestry crane: BC and BCRL deposit 93.8% and 88.9% of pooled inventory against 80.4% for a geometric heuristic, and BCRL succeeds on 83.6% of cycles versus 79.6% for the heuristic and 65.7% for BC, while fine-tuning's simulated stability gain does not carry over to the crane testbed.