The International Journal of Robotics Study, Ahead of Publish.
Model-based support understanding is acknowledged with the possible to be considerably extra example effective than model-free support understanding. Just how an exact design can be created immediately and successfully from raw sensory inputs (such as photos), specifically for complicated settings and jobs, is a difficult issue that prevents the wide application of model-based support understanding in the real life. In this job, we recommend a sensing-aware model-based support discovering system called SAM-RL. Leveraging the differentiable physics-based simulation and making, SAM-RL immediately updates the design by contrasting provided photos with actual raw photos and generates the plan successfully. With the sensing-aware discovering pipe, SAM-RL permits a robotic to choose an interesting perspective to keep an eye on the job procedure. We use our structure to real life experiments for completing 3 control jobs: robot setting up, device control, and deformable item control. We show the efficiency of SAM-RL through comprehensive experiments. Video clips are readily available on our task website.
发布者：Jun Lv，转转请注明出处：https://robotalks.cn/sam-rl-sensing-aware-model-based-reinforcement-learning-via-differentiable-physics-based-simulation-and-rendering/
