Computer Vision News - December 2020

9 Experiment 2: This experiment investigates the effect of mixing real data and simulated data to train the grasping model, showing that the performance is hugely improved by including simulations using simulation-to-real methods. And specifically, for the first setup (grasping of unseen objects), even with a large available dataset of 580,000 real world trials, RL-CycleGAN success rate rises from 87% to 94%. Similar results are obtained on the second setup. Experiment 3: In the end, the authors experiment with fine-tuning the grasping models, using on- policy real data. The amount of off-policy real data to train RL-CycleGAN is then reduced to 5,000 for this, and it is found that, in comparison with state-of-the-art methods based only on on-policy data, it reaches the same performance using less episodes and no domain randomization. Conclusion: Tested on two different robot grasping setups, RL-CycleGAN achieves incredible performance. This work shows significant improvement on real world vision-based robotics. You can just have a look at the examples of robot grasping performed by RL-CycleGAN and make up your own mind on this interesting work. RL-CycleGAN We are closing this year’s reviews on research with this exciting insight on robotics and deep learning. We look forward to yet many more stimulating papers in 2021!