20171207 domain-adaptation
?
2 likes?675 views
This document introduces a study that uses simulated data and domain adaptation techniques to efficiently train policies for robotic grasping. The study trains a neural network policy for grasping using large amounts of simulated data. It then adapts the features and visual appearance of the simulated data to match the real world domain using domain adversarial training and a GAN. In experiments, it finds that domain adaptation improves the grasping policy trained on simulated data, leading to more successful real world grasps.
![Confidential
センスタイムジャパンDA in feature space
? Domain-adversarial neural net (DANN) [3]
7](https://image.slidesharecdn.com/20171207-domain-adaptation-171207114508/85/20171207-domain-adaptation-7-320.jpg)
![Confidential
センスタイムジャパンReferences
13
[1] Bousmalis, K., Irpan, A., Wohlhart, P., Bai, Y., Kelcey, M.,
Kalakrishnan, M., ... & Levine, S. (2017). Using simulation and
domain adaptation to improve efficiency of deep robotic
grasping. arXiv preprint arXiv:1709.07857.
[2] Levine, S., Pastor, P., Krizhevsky, A., Ibarz, J., & Quillen, D.
(2016). Learning hand-eye coordination for robotic grasping with
deep learning and large-scale data collection. The International
Journal of Robotics Research, 0278364917710318.
[3] Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H.,
Laviolette, F., ... & Lempitsky, V. (2016). Domain-adversarial training
of neural networks. Journal of Machine Learning Research, 17(59),
1-35.](https://image.slidesharecdn.com/20171207-domain-adaptation-171207114508/85/20171207-domain-adaptation-13-320.jpg)
20171207 domain-adaptation
- 2. Confidential センスタイムジャパンAbout me ? Name: Taku Yoshioka ? Interests: Bayesian inference, machine learning, deep learning and robotics ? Robot and ROS: 9 months ? Affiliation: SenseTime Japan ′ Computer vision and deep learning ′ https://www.sensetime.jp ′ https://blog.sensetime.jp (lunch blog) ′ Kyoto, Tokyo ′ We are hiring! 2
- 3. Confidential センスタイムジャパンPrevious study 3 ? Learn visual servoing policy with a large-scale data obtained with 14 manipulators ? Train CNN for evaluation of success probability of grasping (target)
- 5. Confidential センスタイムジャパンIdea of the study ? Using simulated environment to learn visual servoing policy for grasping in real environment ? Make simulated features close to real ones by domain adaptation (DA) on: ′ Feature vector in the evaluation network ′ Visual appearance of the simulated environment 5 Feature vector Visual appearance
- 6. Confidential センスタイムジャパン ? Fitting distributions based on random samples, without likelihood function (implicit learning) ? Adversarial training (e.g., GAN for generative model) 6 Fitting distributions
- 9. Confidential センスタイムジャパンDA on visual appearance 9 ? GraspGAN: GAN + Label + Semantics (pixel level) GAN Label Semantics
- 10. Confidential センスタイムジャパンExperiment 10 ? 8M simulated data + 94K?9.4M real data ? Sim-only, Real-only, Sim + Real ? Na?ve mix, domain-specific batch normalization (DBN) + visual randomization, DBN + DANN, DBN + DANN + visual randomization, GraspGAN
- 12. Confidential センスタイムジャパン 12 ? Findings: ′ Efficient training by simulated data ′ Improvement of policy by visual appearance DA ? Future study ′ Consider physical discrepancy ′ Other sensor (e.g., depth) ′ Further improvement of success rate of the task
- 13. Confidential センスタイムジャパンReferences 13 [1] Bousmalis, K., Irpan, A., Wohlhart, P., Bai, Y., Kelcey, M., Kalakrishnan, M., ... & Levine, S. (2017). Using simulation and domain adaptation to improve efficiency of deep robotic grasping. arXiv preprint arXiv:1709.07857. [2] Levine, S., Pastor, P., Krizhevsky, A., Ibarz, J., & Quillen, D. (2016). Learning hand-eye coordination for robotic grasping with deep learning and large-scale data collection. The International Journal of Robotics Research, 0278364917710318. [3] Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., ... & Lempitsky, V. (2016). Domain-adversarial training of neural networks. Journal of Machine Learning Research, 17(59), 1-35.