Robotics: Science and Systems XV 2019
DOI: 10.15607/rss.2019.xv.055
|View full text |Cite
|
Sign up to set email alerts
|

DIViS: Domain Invariant Visual Servoing for Collision-Free Goal Reaching

Abstract: Figure 1: Domain Invariant Visual Servoing (DIViS) learns collision-free goal reaching entirely in simulation using dense multi-step rollouts and a recurrent fully convolutional neural network (bottom). DIViS can directly be deployed on real physical robots with RGB cameras for servoing to visually indicated goals as well as semantic object categories (top). AbstractRobots should understand both semantics and physics to be functional in the real world. While robot platforms provide means for interacting with t… Show more

Help me understand this report

Search citation statements

Order By: Relevance

Paper Sections

Select...
3
1
1

Citation Types

0
12
0

Year Published

2020
2020
2023
2023

Publication Types

Select...
5
2
1

Relationship

0
8

Authors

Journals

citations
Cited by 22 publications
(12 citation statements)
references
References 72 publications
(108 reference statements)
0
12
0
Order By: Relevance
“…RGB-based methods such as [22], [23] do not explicitly consider dynamic environments in their formulation, resulting in purely reactive policies. More specifically, the authors in [8], [24] use a network to estimate whether there is an obstacle 1 meter ahead of the robot from monocular RGB images. As a result, an obstacle moving fast towards the robot from a distance will only be detected when it is close to the robot, and the robot might not have time to react.…”
Section: B Related Workmentioning
confidence: 99%
See 1 more Smart Citation
“…RGB-based methods such as [22], [23] do not explicitly consider dynamic environments in their formulation, resulting in purely reactive policies. More specifically, the authors in [8], [24] use a network to estimate whether there is an obstacle 1 meter ahead of the robot from monocular RGB images. As a result, an obstacle moving fast towards the robot from a distance will only be detected when it is close to the robot, and the robot might not have time to react.…”
Section: B Related Workmentioning
confidence: 99%
“…Also, using depth images instead of RGB in a simulation-based learning setting results in a smaller reality gap and reduces the complexity of the data collection pipeline. In comparison, the authors in [8], [24] use RGB images and had to design 24 worlds in game engines manually, use 200 textures, and 21 different furniture items. Here, we simply use four furniture items and place them randomly across the room.…”
Section: B Related Workmentioning
confidence: 99%
“…[12] learned to translate real and simulation data into a canonical representation to be used as robot observation. Several past works incorporated real images as a complementary source of data for learning control policies inside simulation that can be transferred to the real world [4,27,25]. In contrast to all these works, we generate the simulation environment by recomposing the 3D scene and incorporating simulation physics environment as a computational model for predicting the outcome of robot action trajectories in a closed loop.…”
Section: Related Workmentioning
confidence: 99%
“…Executing a complex task from pixel inputs only is considered remarkable and has been focused on by many studies. For example, visual servoing has been widely implemented in many applications associated with robotic tasks [153] [234]. In today's fast-paced world, an increasing number of researchers have examined deep learning to execute complex behaviors from pixels.…”
Section: Vision-based Robotic Graspmentioning
confidence: 99%