A Fog Robotic System for Dynamic Visual Servoing

Tian, Nan; Tanwani, Ajay Kummar; Chen, Jinfa; Ma, Mas; Zhang, Robert; Huang, Boji; Goldberg, Ken; Sojoudi, Somayeh

doi:10.1109/icra.2019.8793600

“…In a related strand, researchers have used keypoint detection for human pose estimation [30][31][32][33][34]. Nevertheless, in robotics applications, it is not uncommon for objects to be detected via fiducial markers [35][36][37]. Closely related to our work, Deng et al [16] used a particle filter-based approach to predicting the distribution of possible object poses.…”

Section: A Object Pose Estimationmentioning

confidence: 81%

Fast Uncertainty Quantification for Deep Object Pose Estimation

Shi¹,

Zhu²,

Tremblay³

et al. 2020

Preprint

0

View full text Add to dashboard Cite

Deep learning-based object pose estimators are often unreliable and overconfident especially when the input image is outside the training domain, for instance, with sim2real transfer. Efficient and robust uncertainty quantification (UQ) in pose estimators is critically needed in many robotic tasks. In this work, we propose a simple, efficient, and plug-and-play UQ method for 6-DoF object pose estimation. We ensemble 2-3 pre-trained models with different neural network architectures and/or training data sources, and compute their average pairwise disagreement against one another to obtain the uncertainty quantification. We propose four disagreement metrics, including a learned metric, and show that the average distance (ADD) is the best learning-free metric and it is only slightly worse than the learned metric, which requires labeled target data. Our method has several advantages compared to the prior art: 1) our method does not require any modification of the training process or the model inputs; and 2) it needs only one forward pass for each model. We evaluate the proposed UQ method on three tasks where our uncertainty quantification yields much stronger correlations with pose estimation errors than the baselines. Moreover, in a real robot grasping task, our method increases the grasping success rate from 35% to 90%. Video and code are available at https://sites.google.com/view/fastuq.

show abstract

“…In a related strand, researchers have used keypoint detection for human pose estimation [30][31][32][33][34]. Nevertheless, in robotics applications, it is not uncommon for objects to be detected via fiducial markers [35][36][37]. Closely related to our work, Deng et al [16] used a particle filter-based approach to predicting the distribution of possible object poses.…”

Section: A Object Pose Estimationmentioning

confidence: 87%

Fast Uncertainty Quantification for Deep Object Pose Estimation

Shi

¹

,

Zhu

²

,

Tremblay³

et al. 2021

2021 IEEE International Conference on Robotics and Automation (ICRA)

View full text Add to dashboard Cite

Deep learning-based object pose estimators are often unreliable and overconfident especially when the input image is outside the training domain, for instance, with sim2real transfer. Efficient and robust uncertainty quantification (UQ) in pose estimators is critically needed in many robotic tasks. In this work, we propose a simple, efficient, and plug-and-play UQ method for 6-DoF object pose estimation. We ensemble 2-3 pre-trained models with different neural network architectures and/or training data sources, and compute their average pairwise disagreement against one another to obtain the uncertainty quantification. We propose four disagreement metrics, including a learned metric, and show that the average distance (ADD) is the best learning-free metric and it is only slightly worse than the learned metric, which requires labeled target data. Our method has several advantages compared to the prior art: 1) our method does not require any modification of the training process or the model inputs; and 2) it needs only one forward pass for each model. We evaluate the proposed UQ method on three tasks where our uncertainty quantification yields much stronger correlations with pose estimation errors than the baselines. Moreover, in a real robot grasping task, our method increases the grasping success rate from 35% to 90%. Video and code are available at https://sites.google.com/view/fastuq.

show abstract

“…Using shortrange wireless communications technologies such as Zigbee, Bluetooth or Wi-Fi direct, robots get services from the edge resources [134], [145]. A heartbeat protocol is implemented in [206] to deal with network latency of multi-agent cloud robotics. Network bandwidth usage can be substantially preserved by processing the data on edge resources [173], [207].…”

Section: A Resource Poolingmentioning

confidence: 99%

Resource Allocation and Service Provisioning in Multi-Agent Cloud Robotics: A Comprehensive Survey

Afrin

¹

,

Jin

²

,

Rahman

³

et al. 2021

IEEE Commun. Surv. Tutorials

View full text Add to dashboard Cite

Robotic applications nowadays are widely adopted to enhance operational automation and performance of real-world Cyber-Physical Systems (CPSs) including Industry 4.0, agriculture, healthcare, and disaster management. These applications are composed of latency-sensitive, data-heavy, and compute-intensive tasks. The robots, however, are constrained in the computational power and storage capacity. The concept of multi-agent cloud robotics enables robotto-robot cooperation and creates a complementary environment for the robots in executing large-scale applications with the capability to utilize the edge and cloud resources. However, in such a collaborative environment, the optimal resource allocation for robotic tasks is challenging to achieve. Heterogeneous energy consumption rates and application of execution costs associated with the robots and computing instances make it even more complex. In addition, the data transmission delay between local robots, edge nodes, and cloud data centres adversely affects the real-time interactions and impedes service performance guarantee. Taking all these issues into account, this paper comprehensively surveys the state-of-the-art on resource allocation and service provisioning in multi-agent cloud robotics. The paper presents the application domains of multi-agent cloud robotics through explicit comparison with the contemporary computing paradigms and identifies the specific research challenges. A complete taxonomy on resource allocation is presented for the first time, together with the discussion of resource pooling, computation offloading, and task scheduling for efficient service provisioning. Furthermore, we highlight the research gaps from the learned lessons, and present future directions deemed beneficial to further advance this emerging field.

show abstract

A Fog Robotic System for Dynamic Visual Servoing

Cited by 36 publications

References 34 publications

Fast Uncertainty Quantification for Deep Object Pose Estimation

Fast Uncertainty Quantification for Deep Object Pose Estimation

Fast Uncertainty Quantification for Deep Object Pose Estimation

Resource Allocation and Service Provisioning in Multi-Agent Cloud Robotics: A Comprehensive Survey

Contact Info

Product

Resources

About