Web Based GPU Acceleration in Embodied Agent Training Workflow

The deep learning training and generation of virtual scenes, reconstruction or manipulation in the realistic simulation environment are highly compute-intensive tasks. GPU accelerated training of embodied agents inside the realistic 3D scenes in continuous development and improvement process for goal-directed navigation tasks are considered in paper. Web client side computation for reinforcement learning based embodied agent training workflow steps and cloud data delivering are described. Approach with web based GPU interaction technologies (WebGL, WebGPU) along with WebAssembly packed logic (WASM) is proposed. Embodied artificial intelligence (AI) simulation platforms supporting reinforcement learning with realistic assets or powerful logic are tested. The paper presents results that could be used in cases when the training server is deployed in a headless cloud host or in the case when the client machine power utilization is preferable.