A survey of deep learning techniques for autonomous driving

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The majority of the surveyed work is dated after 2005.

“since the majority of the surveyed work is dated after 2005.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Although DRL systems are able to explore different driving situations within a simulated world, these models tend to have a biased behavior when ported to the real-world.

“On the other hand, although DRL systems are able to explore different driving situations within a simulated world, these models tend to have a biased behavior when ported to the real-world.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The evaluation of the Tesla Autopilot comparison framework revealed an accuracy of 90.4% in detecting differences between both systems and the control transfer of the car to a human driver.

“The evaluation revealed an accuracy of 90 . 4% in detecting differences between both systems and the control transfer of the car to a human driver.”

• cites [120] [120] L. Fridman, D. E. Brown, M. Glazer, W. Angell, S. Dodd, B. Jenik, J. Terwilliger, J. Kindelsberger, L. Ding, S. Seaman, H. Abraham, A. Mehler, A. Sipperley, A. Pettinato, L. Angell, B. Mehler, and B. Reimer, 'MIT Autonomous Vehicle Technology Study: Large-Scale Deep Learning Based Analysis of Driver Behavior and Interaction with Automation,' IEEE Access 2017 , 2017. [Online]. Available: https://arxiv.org/abs/1711.06976 -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Imitation Learning has the advantage that it can be trained with data collected from the real-world, but this data is scarce on corner cases, making the trained network's response uncertain when confronted with unseen data.

“IL has the advantage that it can be trained with data collected from the real-world. Nevertheless, this data is scarce on corner cases (e.g. driving off-lanes, vehicle crashes, etc.), making the trained network's response uncertain when confronted with unseen data.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Single stage detectors do not provide the same performances as double stage detectors, but are significantly faster.

“In general, single stage detectors do not provide the same performances as double stage detectors, but are significantly faster.”

May combine two claims.

VO techniques are still dominated by classical keypoints matching algorithms, combined with acceleration data provided by inertial sensors, mainly because keypoints detectors are computationally efficient and can be easily deployed on embedded devices.

“Although much progress has been reported in the area of deep learning based localization, VO techniques are still dominated by classical keypoints matching algorithms, combined with acceleration data provided by inertial sensors. This is mainly due to the fact that keypoints detectors are computational efficient and can be easily deployed on embedded devices.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

PilotNet has 250,000 parameters and approximately 27 million connections.

“PilotNet has 250 . 000 parameters and approx. 27 mil . connections.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The sensors setup differs from vehicle to vehicle, depending on how the data is planned to be used.

“The sensors setup differs from vehicle to vehicle, depending on how the data is planned to be used.”

Classified as probably not a substantial claim.

The authors have identified seven major areas that form open challenges in autonomous driving.

“We have identified seven major areas that form open challenges in the field of autonomous driving.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Many driving datasets were made public and documented in recent years due to increasing research interest in autonomous vehicles.

“In the last years, mainly due to the large and increasing research interest in autonomous vehicles, many driving datasets were made public and documented.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Public driving datasets vary in size, sensor setup and data format.

“They vary in size, sensor setup and data format.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Given the advancements in deep reinforcement learning, the direct application of the algorithm still needs a training pipeline in which one should simulate and model the desired self-driving car's behavior.

“Given the advancements in deep reinforcement learning, the direct application of the algorithm still needs a training pipeline in which one should simulate and model the desired self-driving car's behavior.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The model-based policy iteration algorithm was introduced based on the proof that the Bellman equation is a contraction mapping.

“The model-based policy iteration algorithm was introduced in [18], based on the proof that the Bellman equation is a contraction mapping [20] when written as an operator n :”

• cites [18] [18] R. Sutton and A. Barto, Introduction to Reinforcement Learning . MIT Press, 1998. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [20] [20] C. Watkins and P. Dayan, 'Q-Learning,' Machine Learning , vol. 8, no. 3, p. 279292, 1992. - sourceNot checked yet

  Not checked in the free run.

Classified as probably not a substantial claim. May combine two claims.

The driving decisions in AI-powered autonomous cars are computed either in a modular perception-planning-action pipeline or in an End2End learning fashion, where sensory information is directly mapped to control outputs.

“The driving decisions are computed either in a modular perception-planning-action pipeline (Fig. 1(a)), or in an End2End learning fashion (Fig. 1(b)), where sensory information is directly mapped to control outputs.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The components of the modular pipeline can be designed either based on AI and deep learning methodologies, or using classical non-learning approaches.

“The components of the modular pipeline can be designed either based on AI and deep learning methodologies, or using classical non-learning approaches.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The method in [24] describes a model-free actor-critic algorithm able to learn different continuous control tasks directly from raw pixel inputs.

“The method in [24] describes a model-free actor-critic algorithm able to learn different continuous control tasks directly from raw pixel inputs.”

• cites [24] [24] T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y. Tassa, D. Silver, and D. Wierstra, 'Continuous Control with Deep Reinforcement Learning,' 2-4 May 2016. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

In comparison to traditional handcrafted features, convolutional neural networks are able to automatically learn a representation of the feature space encoded in the training set.

“In comparison to these traditional handcrafted features, convolutional neural networks are able to automatically learn a representation of the feature space encoded in the training set.”

Probably not a substantial claim: may not pass S3 (load-bearing).

A CNN mimics the functioning of the ventral flux, in which different areas of the brain are sensible to specific features in the visual field.

“A CNN mimics the functioning of the ventral flux, in which different areas of the brain are sensible to specific features in the visual field.”

Probably not a substantial claim: may not pass S4 (based).

The standard reinforcement learning method described above is not feasible in high dimensional state spaces.

“However, the standard reinforcement learning method described above is not feasible in high dimensional state spaces.”

Probably not a substantial claim: may not pass S3 (load-bearing).

LSTMs solve the vanishing gradient problem by incorporating three gates, which control the input, output and memory state.

“As opposed to traditional recurrent neural networks, LSTMs solve the vanishing gradient problem by incorporating three gates, which control the input, output and memory state.”

• cites [17] [17] S. Hochreiter and J. Schmidhuber, 'Long Short-term Memory,' Neural computation , vol. 9, no. 8, pp. 1735-1780, 1997. - sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Although continuous control with DRL is possible, the most common strategy for DRL in autonomous driving is based on discrete control.

“Although continuous control with DRL is possible, the most common strategy for DRL in autonomous driving is based on discrete control [26].”

• cites [26] [26] M. Jaritz, R. de Charette, M. Toromanoff, E. Perot, and F. Nashashibi, 'End-to-End Race Driving with Deep Reinforcement Learning,' 2018 IEEE Int. Conf. on Robotics and Automation (ICRA) , pp. 2070-2075, 2018. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Long Short-Term Memory (LSTM) networks are non-linear function approximators for estimating temporal dependencies in sequence data.

“Long Short-Term Memory (LSTM) [17] networks are non-linear function approximators for estimating temporal dependencies in sequence data.”

• cites [17] [17] S. Hochreiter and J. Schmidhuber, 'Long Short-term Memory,' Neural computation , vol. 9, no. 8, pp. 1735-1780, 1997. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

DeepMind's combined algorithm, entitled Rainbow, was able to outperform the independently competing methods.

“A study on how to combine these improvements on deep reinforcement learning has been provided by DeepMind in [22], where the combined algorithm, entitled Rainbow , was able to outperform the independently competing methods.”

• cites [22] [22] M. Hessel, J. Modayil, H. van Hasselt, T. Schaul, G. Ostrovski, W. Dabney, D. Horgan, B. Piot, M. Azar, and D. Silver, 'Rainbow: Combining Improvements in Deep Reinforcement Learning,' 2017. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

A safety monitor is designed to assure the safety of each module in the modular pipeline.

“A safety monitor is designed to assure the safety of each module.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Convolutional Neural Networks are mainly used for processing spatial information, such as images, and can be viewed as image feature extractors and universal non-linear function approximators.

“Convolutional Neural Networks (CNN) are mainly used for processing spatial information, such as images, and can be viewed as image features extractors and universal non-linear function approximators [7], [8].”

• cites [7] [7] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner, 'Gradient-based Learning Applied to Document Recognition,' Proceedings of the IEEE , vol. 86, no. 11, pp. 2278-2324, Nov 1998. -Not checked yet

  Not checked in the free run.

• cites [8] [8] Y. Bengio, A. Courville, and P. Vincent, 'Representation Learning: A Review and New Perspectives,' IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 35, no. 8, pp. 1798-1828, Aug 2013. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

All of the above complementary improvements have been tested on the Atari 2600 challenge.

“All of the above complementary improvements have been tested on the Atari 2600 challenge.”

• cites [22] [22] M. Hessel, J. Modayil, H. van Hasselt, T. Schaul, G. Ostrovski, W. Dabney, D. Horgan, B. Piot, M. Azar, and D. Silver, 'Rainbow: Combining Improvements in Deep Reinforcement Learning,' 2017. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

DeepMind proposes six extensions to the base DQN, each addressing a distinct concern.

“DeepMind [22] proposes six extensions to the base DQN, each addressing a distinct concern:”

• cites [22] [22] M. Hessel, J. Modayil, H. van Hasselt, T. Schaul, G. Ostrovski, W. Dabney, D. Horgan, B. Piot, M. Azar, and D. Silver, 'Rainbow: Combining Improvements in Deep Reinforcement Learning,' 2017. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Before the rise of deep learning, computer vision systems used to be implemented based on handcrafted features, such as HAAR, LBP, or HoG.

“Before the rise of deep learning, computer vision systems used to be implemented based on handcrafted features, such as HAAR [9], Local Binary Patterns (LBP) [10], or Histograms of Oriented Gradients (HoG) [11].”

• cites [9] [9] P. A. Viola and M. J. Jones, 'Rapid Object Detection using a Boosted Cascade of Simple Features,' in 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2001), with CDROM, 8-14 December 2001, Kauai, HI, USA , 2001, pp. 511-518. -Not checked yet

  Not checked in the free run.

• cites [10] [10] T. Ojala, M. Pietik¨ ainen, and D. Harwood, 'A Comparative Study of Texture Measures with Classification Based on Featured Distributions,' Pattern Recognition , vol. 29, no. 1, pp. 51-59, Jan. 1996. -Not checked yet

  Not checked in the free run.

• cites [11] [11] N. Dalal and B. Triggs, 'Histograms of Oriented Gradients for Human Detection,' in In CVPR , 2005, pp. 886-893. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

When compared to a conventional neural network, the only difference is that the learned weights in each unfolded copy of the RNN are averaged, thus enabling the network to share the same weights over time.

“When compared to a conventional neural network, the only difference is that the learned weights in each unfolded copy of the network are averaged, thus enabling the network to shared the same weights over time.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The turning point in self-driving car development acknowledged Machine Learning and AI as central components of autonomous driving.

“This was a turning point in selfdriving cars development, acknowledging Machine Learning and AI as central components of autonomous driving.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The Rectified Linear Unit (ReLU) is the most commonly used activation function in computer vision applications.

“Typically, the Rectified Linear Unit (ReLU) is the most commonly used activation function in computer vision applications [1].”

• cites [1] [1] A. Krizhevsky, I. Sutskever, and G. E. Hinton, 'ImageNet Classification with Deep Convolutional Neural Networks,' in Advances in Neural Information Processing Systems 25 , F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger, Eds. Curran Associates, Inc., 2012, pp. 1097-1105. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

A model-based solution for continuous Q-learning is proposed in [25].

“A model-based solution for continuous Qlearning is proposed in [25].”

• cites [25] [25] S. Gu, T. Lillicrap, I. Sutskever, and S. Levine, 'Continuous Deep Q-Learning with Model-based Acceleration,' in Int. Conf. on Machine Learning ICML 2016 , vol. 48, Jun. 2016, pp. 2829-2838. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In literature, such a non-linear approximator is called a Deep Q-Network (DQN) and is used for estimating the approximate action-value function.

“In literature, such a non-linear approximator is called a Deep QNetwork (DQN) [21] and is used for estimating the approximate action-value function:”

• cites [21] [21] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski, S. Petersen, C. Beattie, A. Sadik, I. Antonoglou, H. King, D. Kumaran, D. Wierstra, S. Legg, and D. Hassabis, 'Human-level Control Through Deep Reinforcement Learning,' Nature , vol. 518, no. 7540, pp. 529-533, Feb. 2015. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The deep reinforcement learning community has made several independent improvements to the original DQN algorithm.

“The deep reinforcement learning community has made several independent improvements to the original DQN algorithm [21].”

• cites [21] [21] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski, S. Petersen, C. Beattie, A. Sadik, I. Antonoglou, H. King, D. Kumaran, D. Wierstra, S. Legg, and D. Hassabis, 'Human-level Control Through Deep Reinforcement Learning,' Nature , vol. 518, no. 7540, pp. 529-533, Feb. 2015. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The visual information is processed according to the dual flux theory, which states that the visual flow follows two main fluxes: a ventral flux responsible for visual identification and object recognition, and a dorsal flux used for establishing spatial relations between objects.

“The information is processed according to the dual flux theory [13], which states that the visual flow follows two main fluxes: a ventral flux , responsible for visual identification and object recognition, and a dorsal flux used for establishing spatial relations between objects.”

• cites [13] [13] M. A. Goodale and A. Milner, 'Separate Visual Pathways for Perception and Action,' Trends in Neurosciences , vol. 15, no. 1, pp. 20 - 25, 1992. - sourceNot checked yet

  Not checked in the free run.

May combine two claims.

In order to decode the true environment state, it is not sufficient to map a single snapshot of sensors readings; the temporal information should also be included in the network's input, since the environment's state is modified over time.

“In order to decode the true environment state, it is not sufficient to map a single snapshot of sensors readings. The temporal information should also be included in the network's input, since the environment's state is modified over time.”

• AssumptionRests on an assumption: The simulated environment state is not directly accessible to the agent.

Probably not a substantial claim: may not pass S3 (load-bearing).

CNNs can be loosely understood as very approximate analogies to different parts of the mammalian visual cortex.

“CNNs can be loosely understood as very approximate analogies to different parts of the mammalian visual cortex [12].”

• cites [12] [12] D. H. Hubel and T. N.Wiesel, 'Shape and Arrangement of Columns in Cats Striate Cortex,' The Journal of Physiology , vol. 165, no. 3, p. 559568, 1963. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The optimal action-value function Q* satisfies the Bellman optimality equation.

“The optimal action-value function Q ∗ satisfies the Bellman optimality equation [19], which is a recursive formulation of Eq. 10:”

• cites [19] [19] R. Bellman, Dynamic Programming . Princeton University Press, 1957. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The main challenge in using basic RNNs is the vanishing gradient encountered during training.

“The main challenge in using basic RNNs is the vanishing gradient encountered during training.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The DQN approach has been extended to continuous action spaces based on policy gradient estimation.

“The DQN approach described above has been extended to continuous action spaces based on policy gradient estimation [24].”

• cites [24] [24] T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y. Tassa, D. Silver, and D. Wierstra, 'Continuous Control with Deep Reinforcement Learning,' 2-4 May 2016. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

In the case of an autonomous car, the discrete actions would translate to discrete commands, such as turn left, turn right, accelerate, or break.

“In the case of an autonomous car, the discrete actions would translate to discrete commands, such as turn left, turn right, accelerate, or break.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Various permutations of learning and non-learning based components are possible in the modular pipeline.

“Various permutations of learning and non-learning based components are possible (e.g. a deep learning based object detector provides input to a classical A-star path planning algorithm).”

Probably not a substantial claim: may not pass S3 (load-bearing).

Deep Learning and AI became the main technologies behind many breakthroughs in computer vision, robotics and NLP over the last decade.

“Over the course of the last decade, Deep Learning and Artificial Intelligence (AI) became the main technologies behind many breakthroughs in computer vision [1], robotics [2] and Natural Language Processing (NLP) [3].”

• cites [1] [1] A. Krizhevsky, I. Sutskever, and G. E. Hinton, 'ImageNet Classification with Deep Convolutional Neural Networks,' in Advances in Neural Information Processing Systems 25 , F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger, Eds. Curran Associates, Inc., 2012, pp. 1097-1105. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [2] [2] M. Andrychowicz, B. Baker, M. Chociej, R. Jozefowicz, B. McGrew, J. Pachocki, A. Petron, M. Plappert, G. Powell, A. Ray, J. Schneider, S. Sidor, J. Tobin, P. Welinder, L. Weng, and W. Zaremba, 'Learning Dexterous In-Hand Manipulation,' CoRR , vol. abs/1808.00177, August 2018. [Online]. Available: https://arxiv.org/abs/1808.00177 - sourceNot checked yet

  Not checked in the free run.

• cites [3] [3] Y. Goldberg, Neural Network Methods for Natural Language Processing , ser. Synthesis Lectures on Human Language Technologies. Morgan & Claypool, 2017, vol. 37. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The success of image based 3D estimation is of high importance to the large scale deployment of autonomous cars, since the LiDAR is arguably one of the most expensive hardware component in a self-driving vehicle.

“The success of image based 3D estimation is of high importance to the large scale deployment of autonomous cars, since the LiDAR is arguably one of the most expensive hardware component in a self-driving vehicle.”

• AssumptionRests on an assumption: LiDAR is arguably one of the most expensive hardware components in a self-driving vehicle.

Probably not a substantial claim: may not pass S3 (load-bearing).

Researchers at Cornell University tried to replicate LiDAR-like point clouds from visual depth estimation.

“Researchers at Cornell University tried to replicate LiDAR-like point clouds from visual depth estimation [32].”

• cites [32] [32] Y. Wang, W.-L. Chao, D. Garg, B. Hariharan, M. Campbell, and K. Weinberger, 'Pseudo-LiDAR from Visual Depth Estimation: Bridging the Gap in 3D Object Detection for Autonomous Driving,' in IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) 2019 , June 2019. -Not checked yet

  Not checked in the free run.

May combine two claims.

AVOD leverages LiDAR data only for training, while images are used during training and deployment, and at deployment stage, AVOD is able to predict 3D bounding boxes of objects solely from image data.

“A solution here would be to use neural network architectures such as AVOD (Aggregate View Object Detection) [34], which leverage on LiDAR data only for training, while images are used during training and deployment. At deployment stage, AVODisable to predict 3D bounding boxes of objects solely from image data.”

• cites [34] [34] J. Ku, M. Mozifian, J. Lee, A. Harakeh, and S. L. Waslander, 'Joint 3D Proposal Generation and Object Detection from View Aggregation,' in IEEE/RSJ Int. Conf. on Intelligent Robots and Systems (IROS) 2018 . IEEE, 2018. -Not checked yet

  Not checked in the free run.

May combine two claims.

End2End methods have been popularized in the last couple of years by NVIDIA, as part of the PilotNet architecture; the approach is to train a CNN which maps raw pixels from a single front-facing camera directly to steering commands.

“End2End methods have been popularized in the last couple of years by NVIDIA ® , as part of the PilotNet architecture. The approach is to train a CNN which maps raw pixels from a single front-facing camera directly to steering commands [119].”

• cites [119] [119] M. Bojarski, D. D. Testa, D. Dworakowski, B. Firner, B. Flepp, P. Goyal, L. D. Jackel, M. Monfort, U. Muller, J. Zhang, X. Zhang, J. Zhao, and K. Zieba, 'End to End Learning for Self-Driving Cars,' CoRR , vol. abs/1604.07316, 2016. [Online]. Available: http://arxiv.org/abs/1604.07316 -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Tesla and Waymo, two of the companies leading the development of self-driving technology, have different philosophies with respect to their main perception sensor, as well as regarding the targeted SAE level.

“Tesla ® and Waymo ® , two of the companies leading the development of self-driving technology [30], have different philosophies with respect to their main perception sensor, as well as regarding the targeted SAE level [4].”

• cites [30] [30] S. O'Kane, 'How Tesla and Waymo are Tackling a Major Problem for Self-Driving Cars: Data,' Transportation , 2018. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [4] [4] SAE Committee, 'Taxonomy and Definitions for Terms Related to On-road Motor Vehicle Automated Driving Systems,' 2014. - sourceNot checked yet

  Not checked in the free run.

May combine two claims.

ISO 26262 provides comprehensive requirements for assuring safety but does not address the unique characteristics of deep learning-based software.

“The ISO 26262 standard for functional safety of road vehicles provides a comprehensive set of requirements for assuring safety, but does not address the unique characteristics of deep learning-based software.”

• cites [130] [130] R. Salay, R. Queiroz, and K. Czarnecki, 'An Analysis of ISO 26262: Using Machine Learning Safely in Automotive Software,' CoRR , vol. abs/1709.02435, 2017. [Online]. Available: http://arxiv.org/abs/1709. 02435 -Not checked yet

  Not checked in the free run.

May combine two claims.

Detectors such as SqueezeNet or [55] are optimized to run on embedded hardware, have a smaller neural network architecture, making it possible to detect objects using a reduced number of operations, at the cost of detection accuracy.

“If in-vehicle computation resources are scarce, one can use detectors such as SqueezeNet [40] or [55], which are optimized to run on embedded hardware. These detectors usually have a smaller neural network architecture, making it possible to detect objects using a reduced number of operations, at the cost of detection accuracy.”

• cites [40] [40] F. N. Iandola, S. Han, M. W. Moskewicz, K. Ashraf, W. J. Dally, and K. Keutzer, 'Squeezenet: Alexnetlevel Accuracy with 50x Fewer Parameters and¡ 0.5 Mb Model Size,' arXiv preprint arXiv:1602.07360 , 2016. - sourceNot checked yet

  Not checked in the free run.

• cites [55] [55] J. Li, K. Peng, and C.-C. Chang, 'An Efficient Object Detection Algorithm Based on Compressed Networks,' Symmetry , vol. 10, no. 7, p. 235, 2018. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Waymo makes use of three Lidar sensors, five radars and eight cameras, while Tesla cars are equipped with eight cameras, 12 ultrasonic sensors and one forward-facing radar.

“For example, alongside three Lidar sensors, Waymo also makes use of five radars and eight cameras, while Tesla ® cars are equipped with eights cameras, 12 ultrasonic sensors and one forward-facing radar.”

• cites [4] [4] SAE Committee, 'Taxonomy and Definitions for Terms Related to On-road Motor Vehicle Automated Driving Systems,' 2014. - sourceNot checked yet

  Not checked in the free run.

May combine two claims.

LiDARs have high resolution and precise perception even in the dark, but are vulnerable to bad weather conditions (e.g. heavy rain) and involve moving parts.

“LiDARs have high resolution and precise perception even in the dark, but are vulnerable to bad weather conditions (e.g. heavy rain) [31] and involve moving parts.”

• cites [31] [31] S. Hasirlioglu, A. Kamann, I. Doric, and T. Brandmeier, 'Test Methodology for Rain Influence on Automotive Surround Sensors,' in 2016 IEEE 19th Int. Conf. on Intelligent Transportation Systems (ITSC) , Nov 2016, pp. 2242-2247. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable).

The use of deep learning components in safety critical systems is still an open question.

“Regardless of the above empirical definitions and possible interpretations of safety, the use of deep learning components in safety critical systems is still an open question.”

May combine two claims.

With every gathered batch of training data, the approximation of the true system model becomes more accurate, thus enabling model flexibility, consistent uncertainty estimates and anticipation of repeatable effects and disturbances that cannot be modeled prior to deployment.

“With every gathered batch of training data, the approximation of the true system model becomes more accurate, thus enabling model flexibility, consistent uncertainty estimates and anticipation of repeatable effects and disturbances that cannot be modeled prior to deployment [91].”

• cites [91] [91] C. J. Ostafew, J. Collier, A. P. Schoellig, and T. D. Barfoot, 'Learning-based Nonlinear Model Predictive Control to Improve Vision-based Mobile Robot Path Tracking,' Journal of Field Robotics , vol. 33, no. 1, pp. 133-152, 2015. -Not checked yet

  Not checked in the free run.

May combine two claims.

The main sensing technologies differ in both companies: Tesla tries to leverage on its camera systems, whereas Waymo's driving technology relies more on Lidar sensors.

“The main sensing technologies differ in both companies. Tesla ® tries to leverage on its camera systems, whereas Waymo's driving technology relies more on Lidar sensors 4 .”

• cites [4] [4] SAE Committee, 'Taxonomy and Definitions for Terms Related to On-road Motor Vehicle Automated Driving Systems,' 2014. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

[130] addresses the gap by analyzing where machine learning can impact ISO 26262 and provides recommendations on how to accommodate this impact.

“[130] addresses this gap by analyzing the places where machine learning can impact the standard and provides recommendations on how to accommodate this impact.”

• cites [130] [130] R. Salay, R. Queiroz, and K. Czarnecki, 'An Analysis of ISO 26262: Using Machine Learning Safely in Automotive Software,' CoRR , vol. abs/1709.02435, 2017. [Online]. Available: http://arxiv.org/abs/1709. 02435 -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Occupancy maps represent an in-vehicle virtual environment, integrating perceptual information in a form better suited for path planning and motion control.

“Occupancy maps represent an in-vehicle virtual environment, integrating perceptual information in a form better suited for path planning and motion control.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Demonstrating the safety of a system which is running deep learning techniques depends heavily on the type of technique and the application context.

“Demonstrating the safety of a system which is running deep learning techniques depends heavily on the type of technique and the application context.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Image data is more suited for the object recognition task because it is richer in information.

“Being richer in information, image data is more suited for the object recognition task.”

May combine two claims.

In [44] and [57], the authors proposed approaches to speed up data processing and inference on embedded devices for autonomous driving; both architectures are light networks providing similar results as SegNet, with reduced computation cost.

“In [44] and [57], the authors proposed approaches to speed up data processing and inference on embedded devices for autonomous driving. Both architectures are light networks providing similar results as SegNet, with a reduced computation cost.”

• cites [44] [44] M. Treml, J. A. Arjona-Medina, T. Unterthiner, R. Durgesh, F. Friedmann, P. Schuberth, A. Mayr, M. Heusel, M. Hofmarcher, M. Widrich, B. Nessler, and S. Hochreiter, 'Speeding up Semantic Segmentation for Autonomous Driving,' 2016. -Not checked yet

  Not checked in the free run.

• cites [57] [57] A. Paszke, A. Chaurasia, S. Kim, and E. Culurciello, 'Enet: A Deep Neural Network Architecture for Real-time Semantic Segmentation,' arXiv preprint arXiv:1606.02147 , 2016. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Waymo is building their vehicles directly as Level 5 systems, with currently more than 10 million miles driven autonomously.

“Waymo ® is building their vehicles directly as Level 5 systems, with currently more than 10 million miles driven autonomously 2 .”

• cites [2] [2] M. Andrychowicz, B. Baker, M. Chociej, R. Jozefowicz, B. McGrew, J. Pachocki, A. Petron, M. Plappert, G. Powell, A. Ray, J. Schneider, S. Sidor, J. Tobin, P. Welinder, L. Weng, and W. Zaremba, 'Learning Dexterous In-Hand Manipulation,' CoRR , vol. abs/1808.00177, August 2018. [Online]. Available: https://arxiv.org/abs/1808.00177 -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

A major advantage of learning controllers is that they optimally combine traditional model-based control theory with learning algorithms, making it possible to still use established methodologies for controller design and stability analysis, together with a robust learning component applied at system identification and prediction levels.

“A major advantage of learning controllers is that they optimally combine traditional model-based control theory with learning algorithms. This makes it possible to still use established methodologies for controller design and stability analysis, together with a robust learning component applied at system identification and prediction levels.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Deep learning based perception, in particular Convolutional Neural Networks (CNNs), became the de-facto standard in object detection and recognition, obtaining remarkable results in competitions such as the ImageNet Large Scale Visual Recognition Challenge.

“Deep learning based perception, in particular Convolutional Neural Networks (CNNs), became the de-facto standard in object detection and recognition, obtaining remarkable results in competitions such as the ImageNet Large Scale Visual Recognition Challenge [35].”

• cites [35] [35] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and L. FeiFei, 'ImageNet Large Scale Visual Recognition Challenge,' Int. Journal of Computer Vision (IJCV) , vol. 115, no. 3, pp. 211-252, 2015. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The survey in [82] does not include a review on deep learning technologies, although the state of the art literature has revealed an increased interest in using deep learning technologies for path planning and behavior arbitration.

“However, [82] does not include a review on deep learning technologies, although the state of the art literature has revealed an increased interest in using deep learning technologies for path planning and behavior arbitration.”

• cites [82] [82] S. D. Pendleton, H. Andersen, X. Du, X. Shen, M. Meghjani, Y. H. Eng, D. Rus, and M. H. Ang, 'Perception, Planning, Control, and Coordination for Autonomous Vehicles,' Machines , vol. 5, no. 1, p. 6, 2017. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The Oxford dataset collection spanned over 1 year, resulting in over 1000 km of recorded driving with almost 20 million images from 6 cameras, along with LIDAR, GPS and INS ground truth.

“Provided by Oxford University, UK, the dataset collection spanned over 1 year, resulting in over 1000 km of recorded driving with almost 20 million images collected from 6 cameras mounted to the vehicle, along with LIDAR, GPS and INS ground truth.”

• cites [153] [153] W. Maddern, G. Pascoe, C. Linegar, and P. Newman, '1 Year, 1000km: The Oxford RobotCar Dataset,' The Int. Journal of Robotics Research (IJRR) , vol. 36, no. 1, pp. 3-15, 2017. -Not checked yet

  Not checked in the free run.

May combine two claims.

FPGAs provide a significant advantage over GPUs in developing driver assistance systems because FPGAs are designed to meet safety requirements while GPUs were originally built for graphics and high-performance computing where functional safety is not necessary.

“When compared to GPUs, which were originally built for graphics and high-performance computing systems, where functional safety is not necessary, FPGAs provide a significant advantage in developing driver assistance systems.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The final NuScenes dataset will include approximately 1.4 million camera images, 400,000 Lidar sweeps, 1.3 million RADAR sweeps and 1.1 million object bounding boxes in 40,000 keyframes.

“The final dataset will include approximately 1,4 million camera images, 400.000 Lidar sweeps, 1,3 million RADAR sweeps and 1,1 million object bounding boxes in 40.000 keyframes.”

• cites [146] [146] H. Caesar, V. Bankiti, A. H. Lang, S. Vora, V. E. Liong, Q. Xu, A. Krishnan, Y. Pan, G. Baldan, and O. Beijbom, 'nuScenes: A multimodal Dataset for Autonomous Driving,' arXiv preprint arXiv:1903.11027 , 2019. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

The main limitations for vehicles are price, performance issues and power consumption.

“The main limitations when it comes to vehicles are the price, performance issues and power consumption.”

May combine two claims.

The probability distribution is usually unknown, precluding the use of domain adaptation techniques.

“Usually the probability distribution is unknown, precluding the use of domain adaptation techniques [134] [135].”

• cites [134] [134] H. Daum´ e, III and D. Marcu, 'Domain Adaptation for Statistical Classifiers,' J. Artif. Int. Res. , vol. 26, no. 1, pp. 101-126, May 2006. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [135] [135] R. Caruana, Y. Lou, J. Gehrke, P. Koch, M. Sturm, and N. Elhadad, 'Intelligible Models for HealthCare: Predicting Pneumonia Risk and Hospital 30day Readmission,' in Proceedings of the 21th ACM SIGKDD Int. Conf. on Knowledge Discovery and Data Mining , 2015, pp. 1721-1730. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Machine learning and deep learning techniques are starting to become effective and reliable even for safety critical systems, though complete safety assurance remains an open question.

“Machine learning and deep learning techniques are starting to become effective and reliable even for safety critical systems, even if the complete safety assurance for this type of systems is still an open question.”

May combine two claims.

FPGAs show great improvements in both performance and power consumption in deep learning applications.

“AField-Programmable Gate Array (FPGA) is another viable solution, showing great improvements in both performance and power consumption in deep learning applications.”

• cites [170] sourceNot checked yet

  Not checked in the free run.

• cites [171] sourceNot checked yet

  Not checked in the free run.

• cites [172] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable).

Additional work is required for bridging the gap between image- and LiDAR-based 3D perception.

“Furthermore, additional work is required for bridging the gap between image- and LiDAR-based 3D perception [32], enabling the computer vision community to close the current debate on camera vs. LiDAR as main perception sensors.”

• cites [32] [32] Y. Wang, W.-L. Chao, D. Garg, B. Hariharan, M. Campbell, and K. Weinberger, 'Pseudo-LiDAR from Visual Depth Estimation: Bridging the Gap in 3D Object Detection for Autonomous Driving,' in IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) 2019 , June 2019. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The high amount of data needed in the development stage made data collection on public roads a valuable activity.

“The high amount of data needed in the development stage of such components made data collection on public roads a valuable activity.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The majority of available datasets consider sunny, daylight and urban conditions, which are ideal operating conditions for autonomous driving systems.

“The majority of the available datasets consider sunny, daylight and urban conditions, these being ideal operating conditions for autonomous driving systems.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

NVIDIA DRIVE AGX developer kit platform was presented as the world's most advanced self-driving car platform, based on Volta technology.

“Introduced in September 2018, NVIDIA ® DRIVE AGX developer kit platform was presented as the world's most advanced self-driving car platform [166], being based on the Volta technology [167].”

• cites [166] [166] --, 'NVIDIA Drive AGX,' https://www.nvidia.com/en-us/self-drivingcars/drive-platform/hardware/. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [167] [167] --, 'NVIDIA Volta,' https://www.nvidia.com/enus/data-center/volta-gpu-architecture/. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Deploying deep learning algorithms on target edge devices is not a trivial task.

“Deploying deep learning algorithms on target edge devices is not a trivial task.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

DRIVE AGX Xavier is an energy-efficient computing platform with 30 trillion operations per second, meeting ISO 26262 functional safety specification.

“DRIVE AGX Xavier is a scalable open platform that can serve as an AI brain for self driving vehicles, and is an energy-efficient computing platform, with 30 trillion operations per second, while meeting automotive standards like the ISO 26262 functional safety specification.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

NuScenes contains 1000 driving scenes collected from Boston and Singapore, with 25 object classes annotated with 3D bounding boxes at 2Hz.

“Constructed by nuTonomy, this dataset contains 1000 driving scenes collected from Boston and Singapore, two known for their dense traffic and highly challenging driving situations. In order to facilitate common computer vision tasks, such as object detection and tracking, the providers annotated 25 object classes with accurate 3D bounding boxes at 2Hz over the entire dataset.”

• cites [146] [146] H. Caesar, V. Bankiti, A. H. Lang, S. Vora, V. E. Liong, Q. Xu, A. Krishnan, Y. Pan, G. Baldan, and O. Beijbom, 'nuScenes: A multimodal Dataset for Autonomous Driving,' arXiv preprint arXiv:1903.11027 , 2019. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

NVIDIA DRIVE PX is an AI car computer designed to enable auto-makers to focus directly on software for autonomous vehicles.

“DRIVE PX [162] is an AI car computer which was designed to enable the auto-makers to focus directly on the software for autonomous vehicles.”

• cites [162] [162] NVIDIA, 'NVIDIA AI Car Computer Drive PX,' https://www.nvidia.com/en-au/self-drivingcars/drive-px/. - [169] --, 'R-Car H3,' https://www.renesas.com/sg/en/solutions/automotive/soc/rcar-h3.html/. - [170] E. Nurvitadhi, G. Venkatesh, J. Sim, D. Marr, R. Huang, J. Ong Gee Hock, Y. T. Liew, K. Srivatsan, D. Moss, S. Subhaschandra, and G. Boudoukh, 'Can FPGAs Beat GPUs in Accelerating Next-Generation Deep Neural Networks?' in Proceedings of the 2017 ACM/SIGDA Int. Symposium on Field-Programmable Gate Arrays , ser. FPGA '17. New York, NY, USA: ACM, 2017, pp. 5-14. [Online]. Available: http://do sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The study in [120] involved 99 participants, 29 vehicles, 405,807 miles and approximately 5.5 billion video frames.

“The study involved 99 participants, 29 vehicles, 405 , 807 miles and approximatively 5 . 5 billion video frames.”

• cites [120] [120] L. Fridman, D. E. Brown, M. Glazer, W. Angell, S. Dodd, B. Jenik, J. Terwilliger, J. Kindelsberger, L. Ding, S. Seaman, H. Abraham, A. Mehler, A. Sipperley, A. Pettinato, L. Angell, B. Mehler, and B. Reimer, 'MIT Autonomous Vehicle Technology Study: Large-Scale Deep Learning Based Analysis of Driver Behavior and Interaction with Automation,' IEEE Access 2017 , 2017. [Online]. Available: https://arxiv.org/abs/1711.06976 -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The usage of real world data is a key requirement for training and testing an autonomous driving component.

“Undeniably, the usage of real world data is a key requirement for training and testing an autonomous driving component.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Deep learning is the main technology behind a large number of perception systems.

“Deep learning is the main technology behind a large number of perception systems.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Each dataset heavily depends on the objective of the algorithm for which the data was collected.

“Each dataset heavily depends on the objective of the algorithm for which the data was collected.”

May combine two claims.

Long-term reasoning is solved, as provided by navigation systems, which are standard methods for selecting a route through the road network.

“We consider long-term reasoning as solved, as provided by navigation systems. These are standard methods for selecting a route through the road network, from the car's current position to destination [82].”

• cites [82] [82] S. D. Pendleton, H. Andersen, X. Du, X. Shen, M. Meghjani, Y. H. Eng, D. Rus, and M. H. Ang, 'Perception, Planning, Control, and Coordination for Autonomous Vehicles,' Machines , vol. 5, no. 1, p. 6, 2017. -Not checked yet

  Not checked in the free run.

May combine two claims.

Hazards unique to deep reinforcement learning components can be traced to faults and fit within existing ISO 26262 guidelines.

“Although hazards such as these may be unique to deep reinforcement learning components, they can be traced to faults, thus fitting within the existing guidelines of ISO 26262.”

May combine two claims.

Current perception systems are mainly designed to calculate 2D or 3D bounding boxes for a couple of trained object classes, or to provide a segmented image of the driving environment.

“Although great progress has been reported with respect to accuracy in object detection and recognition [173], current systems are mainly designed to calculate 2D or 3D bounding boxes for a couple of trained object classes, or to provide a segmented image of the driving environment.”

• cites [173] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

ISO 26262 states that component behavior shall be fully specified and each refinement shall be verified with respect to its specification.

“ISO 26262 standard states that the component behavior shall be fully specified and each refinement shall be verified with respect to its specification.”

May combine two claims.

Despite 130 million miles of testing, the Tesla accident was caused under extremely rare circumstances (Black Swans) due to truck height, white color under bright sky, and vehicle positioning.

“Despite the 130 million miles of testing and evaluation, the accident was caused under extremely rare circumstances, also known as Black Swans, given the height of the truck, its white color under bright sky, combined with the positioning of the vehicle across the road.”

• cites [137] [137] S. Levin, 'Tesla Fatal Crash: 'Autopilot' Mode Sped up Car Before Driver Killed, Report Finds,' The Guardian , 2018. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Current standards and regulation from the automotive industry cannot be fully mapped to deep learning systems, requiring new safety standards targeted for deep learning.

“Current standards and regulation from the automotive industry cannot be fully mapped to such systems, requiring the development of new safety standards targeted for deep learning.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

NVIDIA DRIVE PX is capable of real-time environment perception, path planning and localization, combining deep learning, sensor fusion and surround vision.

“NVIDIA ® DRIVE PX is capable to perform real-time environment perception, path planning and localization. It combines deep learning, sensor fusion and surround vision to improve the driving experience.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The newest version of DrivePX architecture is based on two Tegra X2 systems on a chip, each containing two Denver cores, 4 ARM A57 cores and a Pascal-generation GPU.

“The newest version of DrivePX architecture is based on two Tegra X2 [163] systems on a chip (SoCs). Each SoC contains two Denve [164] cores, 4 ARM A57 cores and a graphical computeing unit (GPU) from the Pascal [165] generation.”

• cites [163] [163] --, 'Tegra X2,' https://devblogs.nvidia.com/jetson-tx2-deliverstwice-intelligence-edge/. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [164] [164] --, 'Denver Core,' https://en.wikichip.org/wiki/nvidia/microarchitectures/denver. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [165] [165] --, 'Pascal Microarchitecture,' https://www.nvidia.com/en-us/data-center/pascalgpu-architecture/. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

The data collected in the study in [120] is not available for the public.

“Unfortunately, the data collected in this study is not available for the public.”

• cites [120] [120] L. Fridman, D. E. Brown, M. Glazer, W. Angell, S. Dodd, B. Jenik, J. Terwilliger, J. Kindelsberger, L. Ding, S. Seaman, H. Abraham, A. Mehler, A. Sipperley, A. Pettinato, L. Angell, B. Mehler, and B. Reimer, 'MIT Autonomous Vehicle Technology Study: Large-Scale Deep Learning Based Analysis of Driver Behavior and Interaction with Automation,' IEEE Access 2017 , 2017. [Online]. Available: https://arxiv.org/abs/1711.06976 -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Faults and failures of a programmed component are totally different from those of a deep learning component.

“Faults and failures of a programmed component (e.g. one using a formal algorithm to solve a problem) are totally different from the ones of a deep learning component.”

May combine two claims.

The first autonomous driving accident was produced by a Tesla car, where object misclassification errors caused AutoPilot to collide with a truck.

“We must mention the first autonomous driving accident, produced by a Tesla ® car, where, due to object misclassification errors, the AutoPilot function collided the vehicle into a truck [137].”

• cites [137] [137] S. Levin, 'Tesla Fatal Crash: 'Autopilot' Mode Sped up Car Before Driver Killed, Report Finds,' The Guardian , 2018. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Embedded platforms are becoming essential for integration of AI algorithms inside vehicles due to their portability, versatility, and energy efficiency.

“Therefore, embedded platforms are becoming essential for integration of AI algorithms inside vehicles due to their portability, versatility, and energy efficiency.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

KITTI is known as the most prestigious dataset in the self-driving vehicles domain and counts more than 2000 citations.

“It is known as the most prestigious dataset in the self-driving vehicles domain. To this date it counts more than 2000 citations in the literature.”

• cites [151] [151] A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, 'Vision Meets Robotics: The KITTI Dataset,' The Int. Journal of Robotics Research , vol. 32, no. 11, pp. 12311237, 2013. - sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Specific faults of a deep learning component can be caused by unreliable or noisy sensor signals, neural network topology, learning algorithm, training set, or unexpected changes in the environment.

“Specific faults of a deep learning component can be caused by unreliable or noisy sensor signals (video signal due to bad weather, radar signal due to absorbing construction materials, GPS data, etc.), neural network topology, learning algorithm, training set or unexpected changes in the environment (e.g. unknown driving scenes or accidents on the road).”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

There is still a gap to be filled between the accuracy of a simulated world and real-world driving.

“There is still a gap to be filled between the accuracy of a simulated world and real-world driving.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The effectiveness of deep learning systems is directly tied to the availability of training data.

“The effectiveness of deep learning systems is directly tied to the availability of training data.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The usage of deep learning in safety-critical systems is still an open debate.

“The usage of deep learning in safetycritical systems is still an open debate, efforts being made to bring the computational intelligence and functional safety communities closer to each other.”

May combine two claims.

Real-time requirements have to be fulfilled for processing large amounts of data from the car's sensors suite and for updating parameters of deep learning systems over high-speed communication lines.

“Finally, realtime requirements have to be fulfilled for processing the large amounts of data gathered from the car's sensors suite, as well as for updating the parameters of deep learning systems over high-speed communication lines [170].”

• cites [170] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based).

There are still open issues on the explainability, stability, or classification robustness of deep neural networks.

“Although new data-driven design methodologies have been proposed, there are still opened issues on the explainability, stability, or classification robustness of deep neural networks.”

Probably not a substantial claim: may not pass S4 (based).

Most driving scenarios are considered solvable with classical methodologies.

“Most driving scenarios are considered solvable with classical methodologies.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The better the quality of the data is, the higher the accuracy of the algorithm.

“The better the quality of the data is, the higher the accuracy of the algorithm.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Deep learning systems rely on large training databases and require extensive computational hardware.

“On top of this, deep learning systems rely on large training databases and require extensive computational hardware.”

May combine two claims.

The daily data recorded by an autonomous vehicle poses challenges on the parallelization of the training procedure, as well as on the storage infrastructure.

“This poses challenges on the parallelization of the training procedure, as well as on the storage infrastructure.”

• AssumptionRests on an assumption: The daily data recorded by an autonomous vehicle is on the order of petabytes.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Autonomous vehicle technology has seen rapid progress in the past decade, especially due to advances in AI and deep learning.

“Autonomous vehicle technology has seen a rapid progress in the past decade, especially due to advances in the area of artificial intelligence and deep learning.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

A major challenge is the difficulty in proving the functional safety of autonomous vehicles, given the current formalism and explainability of neural networks.

“A major challenge is the difficulty in proving the functional safety of these vehicle, given the current formalism and explainability of neural networks.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

These real-time constraints can be backed up by advances in semiconductor chips dedicated for self-driving cars, as well as by the rise of 5G communication networks.

“These real-time constraints can be backed up by advances in semiconductor chips dedicated for self-driving cars, as well as by the rise of 5G communication networks.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Current AI methodologies are nowadays either used or taken into consideration when designing different components for a self-driving car.

“Current AI methodologies are nowadays either used or taken into consideration when designing different components for a self-driving car.”

Classified as probably not a substantial claim.

Generalization in deep learning is of special importance in learning hazardous situations that can lead to accidents, especially because training data for such corner cases is scarce.

“Generalization in deep learning is of special importance in learning hazardous situations that can lead to accidents, especially due to the fact that training data for such corner cases is scarce.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Deep learning approaches have influenced the design of traditional perception-planning-action pipelines and enabled End2End learning systems that directly map sensory information to steering commands.

“Deep learning approaches have influenced not only the design of traditional perception-planning-action pipelines, but have also enabled End2End learning systems, able do directly map sensory information to steering commands.”

Classified as probably not a substantial claim.

Classical controllers make use of an a-priori model composed of fixed parameters.

“Classical controllers make use of an a-priori model composed of fixed parameters.”

Probably not a substantial claim: may not pass S4 (based).

The remaining unsolved scenarios are corner cases which, until now, required the reasoning and intelligence of a human driver.

“However, the remaining unsolved scenarios are corner cases which, until now, required the reasoning and intelligence of a human driver.”