Fang, Zhixin2025-04-232025-04-232025https://hdl.handle.net/1828/21965Robotic deep RL tasks are intrinsically hard to parallelize, as they usually require cooperation between the RL agent, the simulator and the physics engine. Context barriers can be hard to cross between those systems, making device-host transfer costs during training a challenge. Furthermore, the rapidly evolving world of machine learning also makes integration increasingly difficult for researchers, as incorporating a new learning algorithm usually requires huge engineering effort. TocabiLab is developed as an extension to NVIDIA Isaac Lab with the capability of conducting GPU-native robotics deep RL tasks. Its unified API inherited from Isaac Lab reduces the difficulty of integration, as well as enabling data to be swiftly passed between components at a large scale, accelerating training under computation intensive scenarios. This project report describes the novel architecture applied in the proposed system and a set of experiments designed upon a previously published research to validate the system’s correctness, efficiency and versatility. Evaluation result shows that the system is capable of handling more than 300% load than the industry standard systems while providing comparable performance.enAvailable to the World Wide Webreinforcement learningGPU programmingroboticshumanoid and bipedal locomotionproject