Using Simulation and Domain Adaptation to Improve Efficiency of Deep Robotic Grasping

Konstantinos Bousmalis, Alex Irpan, Paul Wohlhart, Yunfei Bai, Matthew Kelcey, Mrinal Kalakrishnan, Laura Downs, Julian Ibarz, Peter Pastor, Kurt Konolige, Sergey Levine, Vincent Vanhoucke

Seminal September 2017

TL;DR Train a monocular-RGB grasp-success CNN on 8 million simulated grasps of random procedural shapes plus real data, with GraspGAN making the rendered images look real and a domain-adversarial loss aligning features. With 2% of the real data it matches the all-real model (68.51% vs 67.65%), and with no real labels it reaches 63.40%.

What they do

Learned grasping from monocular RGB needs many labeled real grasps. This paper keeps the grasping system of Levine et al. unchanged (a CNN predicting grasp success from images and a motion command, plus a servoing loop) and studies how simulated data can replace real data: which simulated objects, which scene randomization, and which domain adaptation. The evaluation uses 25,704 real grasps of 36 unseen objects.

How it works

Simulation. A Bullet copy of the Kuka bin setup, with 1,000 to 2,000 simulated arms collecting about 8 million samples. The objects are either 1,000 procedurally generated shapes made of random rectangular prisms or ShapeNet models.
Mixing and feature-level adaptation. Simulated and real samples share each batch, with separate batch-norm statistics per domain (DBN). A DANN domain classifier with gradient reversal sits on the grasp network’s last convolutional layer.
GraspGAN. A U-Net generator restyles simulated images to look real, and a multi-scale patch discriminator scores the two input images together. Content losses keep the scene intact: the generator also predicts segmentation masks, and the task network’s activations must match on the simulated and adapted images.

Results

With 1% of the real data (93,841 samples), success rises from 31.13% (real only) to 59.95% (GraspGAN). With all 9.4 million real samples it rises from 67.65% to 76.67%.
GraspGAN with 2% of the real data reaches 68.51%, matching the all-real model with 50x fewer real samples.
Without any real labels, GraspGAN trained on unlabeled real images reaches 63.40%. Sim-only gets 23.53% and randomized sim gets 35.95%, while real-only with 939,777 labeled samples gets 62.75%.
Procedural shapes beat ShapeNet models in every randomization setting (71.93-74.88% vs 68.62-69.84%). Visual randomization helped only slightly.

Limitations

Both adaptation methods target appearance only, with no explicit reasoning about physical differences between simulation and reality. The system handles only top-down pinch grasps from monocular RGB, and the absolute success rates leave room for improvement. Adding DANN to randomized data gave worse results in 3 of 5 cases.

Videos

Figures

From Figure 1: (a) synthetic images from the simulator, (b) the same images adapted by GraspGAN, and (c) real images ... — From Figure 1: (a) synthetic images from the simulator, (b) the same images adapted by GraspGAN, and (c) real images from the camera over the robot's shoulder.

From Figure 4: (a) GraspGAN: the generator G turns simulated image pairs into realistic ones, the discriminator D tel... — From Figure 4: (a) GraspGAN: the generator G turns simulated image pairs into realistic ones, the discriminator D tells them apart from unlabeled real images, and the grasp success network C trains on adapted and real images in parallel, so G gets feedback from both D and C. (b) Architectures of G and D. (c) The DANN model, a domain classifier with a gradient reversal layer on the features of C.

From Figure 5: real grasp success for sim-only, real-only and sim+real training as the amount of real data grows, usi... — From Figure 5: real grasp success for sim-only, real-only and sim+real training as the amount of real data grows, using 8 million simulated samples of procedural objects with no randomization.

Abstract

Instrumenting and collecting annotated visual grasping datasets to train modern machine learning algorithms can be extremely time-consuming and expensive. An appealing alternative is to use off-the-shelf simulators to render synthetic data for which ground-truth annotations are generated automatically. Unfortunately, models trained purely on simulated data often fail to generalize to the real world. We study how randomized simulated environments and domain adaptation methods can be extended to train a grasping system to grasp novel objects from raw monocular RGB images. We extensively evaluate our approaches with a total of more than 25,000 physical test grasps, studying a range of simulation conditions and domain adaptation methods, including a novel extension of pixel-level domain adaptation that we term the GraspGAN. We show that, by using synthetic data and domain adaptation, we are able to reduce the number of real-world samples needed to achieve a given level of performance by up to 50 times, using only randomly generated simulated objects. We also show that by using only unlabeled real-world data and our GraspGAN methodology, we obtain real-world grasping performance without any real-world labels that is similar to that achieved with 939,777 labeled real-world samples.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong