Towards reinforcement learning based log loading automation

Abstract

Forestry forwarders transport cut logs from the felling site to a landing area or a secondary transport vehicle, a task that is physically and mentally demanding for the operator. Even partial automation of the loading cycle can reduce workload and improve safety. This study extends prior reinforcement-learning (RL) work on grasping to the full pick–lift–transport–deliver sequence. We train an agent with Proximal Policy Optimization (PPO) and a two-stage curriculum in a GPU-accelerated simulator (NVIDIA Isaac Gym), using a trailer-type forwarder model with a hydraulic crane and grapple. The task is simplified to a single log and a fixed target location inside the bunk: the agent must reach a randomly placed log, grasp it, lift it above the bed guards, and deliver it with reduced vertical impact velocity. The reward is decomposed into three components, reaching (r1) lifting/unloading (r2), and stable target delivery (r3), and we compare curriculum compositions. In an evaluation over 1,024 parallel episodes, the best two-stage configuration (r1 + r2, then + r3) reaches a 94% success rate, outperforming flat and fully staged alternatives. Generalization tests on unseen log sizes, elevated ground, and rough terrain show partial transfer with clear degradation under larger shifts. The study is limited to a highly simplified simulation setting (flat ground, fixed forwarder base, single log), and no sim-to-real transfer was attempted.