Computer Vision News - December 2020

8 The total loss of the RL-CycleGAN is ultimately defined as: The RL-CycleGAN is evaluated through three experiments on two different setups for robot grasping. To demonstrate the method invariability to robot and task, the two setups have different purposes. The former aims to generalize grasping of unseen objects, and the second increases the complexity of the environment, since grasping is performed from three bins with the robot placed at different locations relative to the bin. The performance is measured in terms of grasp success rate on the two robotic grasping systems. Experiment 1: This experiment is built to analyze if the main aim of bridging the gap when using simulatedexperience is fulfilled. It compares severalmethodsbuiltonGANapproaches and measures the success rate of each. The grasping model trained with the RL- CycleGAN performs the best ( 70% success ). “It preserves task-salient information and produces realistic images and does so with a general-purpose consistency loss that is based directly on the similarity of Q-values, without requiring manual identification of task-salient properties (e.g., object geometry).” The difference with a regular GAN (with a success rate of 29%) is huge, and this can be fully appreciated visually in the figure below. Research ℒ ( ) = ( , , , ′ ) ( ( , ), + ( ′ ) ℒ − ( , , , , ) = ℒ ( , ) + ℒ ( , ) + ℒ ( , ) + − ℒ − ( , ) + ℒ ( )