JetBot Combines CNNs With Reinforcement Learning

A developer trains an NVIDIA JetBot in Isaac Sim and Isaac Lab using SKRL and PPO to navigate from random starts to target coordinates while avoiding 10 randomized 50kg cuboid obstacles in a 300m×300m simulated plane. The agent uses stacked 64×64 RGB frames, a custom 4‑layer CNN with an MLP policy head, continuous two-wheel velocity actions, and a Direct Workflow to speed parallelized training.