Towards robotic cognition using deep neural network applied in a goalkeeper robot

Silva, Isaac Jesus da; Vilao, Claudio O.; Costa, Anna Helena Reali; Bianchi, Reinaldo A. C.

doi:10.1109/sbr-lars-r.2017.8319463

Cited by 5 publications

(3 citation statements)

References 6 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…However, because the output of the semantic segmentation is in the same form as a lookup table based labelling approach, any already existing methods built on top of such a method can directly be reused. For instance, an efficient-and still task agnostic-connected component based method previously developed by us readily fits onto the architecture outlined here and performs the final object detection step within only 1 to 2 ms. N/A 11-22 Task dependent region proposal Cruz et al [5] 24 × 24 440 Task dependent region proposal Javadi et al [10] N/A 240 no loss: 6 fps; task dependent Da Silva et al [6] 110 × 110 8 Predict end-to-end desired action Hess et al [8] 32 × 32 50 Focus on generation of training data Schnekenburger et al [14] 640 × 512 111 GTX-760 GPU; 19 fps on i7 CPU Ours 640 × 480 5 L3F5M2S2 320 × 256 15 L3F5M2S2; L3F4M1.5S2: 20 fps By delaying the use of task dependent methods, one actually has an opportunity to optimise the segmentation output for such methods, by varying the threshold used to determine the final class pixels. For specific use cases it may be desirable to choose a threshold that represents a preference for either high true positive rate (recall), e.g.…”

Section: Discussionmentioning

confidence: 99%

Deep Learning for Semantic Segmentation on Minimal Hardware

Dijk

Scheunemann

2019

RoboCup 2018: Robot World Cup XXII

View full text Add to dashboard Cite

Deep learning has revolutionised many fields, but it is still challenging to transfer its success to small mobile robots with minimal hardware. Specifically, some work has been done to this effect in the RoboCup humanoid football domain, but results that are performant and efficient and still generally applicable outside of this domain are lacking. We propose an approach conceptually different from those taken previously. It is based on semantic segmentation and does achieve these desired properties. In detail, it is being able to process full VGA images in real-time on a low-power mobile processor. It can further handle multiple image dimensions without retraining, it does not require specific domain knowledge for achieving a high frame rate and it is applicable on a minimal mobile hardware.

show abstract

Section: Discussionmentioning

confidence: 99%

Deep Learning for Semantic Segmentation on Minimal Hardware

Dijk

Scheunemann

2019

RoboCup 2018: Robot World Cup XXII

View full text Add to dashboard Cite

show abstract

“…Therefore, the cognitive operator will be implemented with planning and self-optimizing algorithms. Cloud computing principles, especially for robotics [30,37] (e.g., neural network applications [29]), will be evaluated with the DAEbot. The DAEbot will also be used in a multirobot network [38] (e.g., alongside the AMiRo (Autonomous Mini Robot) [1]) to test the multilayer robot cooperation capabilities of the presented architecture.…”

Section: Discussionmentioning

confidence: 99%

“…Unlike the original OCM approach, our architecture implements the cognitive operator as a cloud service. The benefit of this distribution is that the local part can be equipped with processing units which are low on energy but also do not have the computing power to solve complex software algorithms (e.g., particle filters with a high amount of particles, or neural network applications [29]). The cloud part allows adding this computing power while not using the limited battery power of the mobile robot.…”

Section: Autonomy Vs Cloud Computingmentioning

confidence: 99%

Concepts of a Modular System Architecture for Distributed Robotic Systems

2019

View full text Add to dashboard Cite

Modern robots often use more than one processing unit to solve the requirements in robotics. Robots are frequently designed in a modular manner to fulfill the possibility to be extended for future tasks. The use of multiple processing units leads to a distributed system within one single robot. Therefore, the system architecture is even more important than in single-computer robots. The presented concept of a modular and distributed system architecture was designed for robotic systems. The architecture is based on the Operator-Controller Module (OCM). This article describes the adaption of the distributed OCM for mobile robots considering the requirements on such robots, including, for example, real-time and safety constraints. The presented architecture splits the system hierarchically into a three-layer structure of controllers and operators. The controllers interact directly with all sensors and actuators within the system. For that reason, hard real-time constraints need to comply. The reflective operator, however, processes the information of the controllers, which can be done by model-based principles using state machines. The cognitive operator is used to optimize the system. The article also shows the exemplary design of the DAEbot, a self-developed robot, and discusses the experience of applying these concepts on this robot.

show abstract