Search

This paper proposes a strategy for visual perception in the context of autonomous driving. Humans, when not distracted or drunk, are still the best drivers you can currently find. For this reason, we take inspiration from two theoretical ideas about the human mind and its neural organization. The first idea concerns how the brain uses structures of neuron ensembles that expand and compress information to extract abstract concepts from visual experience and code them into compact representations. The second idea suggests that these neural perceptual representations are not neutral but functional to predicting the future state of affairs in the environment. Similarly, the prediction mechanism is not neutral but oriented to the planning of future action. We identify within the deep learning framework two artificial counterparts of the aforementioned neurocognitive theories. We find a correspondence between the first theoretical idea and the architecture of convolutional autoencoders, while we translate the second theory into a training procedure that learns compact representations which are not neutral but oriented to driving tasks, from two distinct perspectives. From a static perspective, we force separate groups of neural units in the compact representations to represent specific concepts crucial to the driving task distinctly. From a dynamic perspective, we bias the compact representations to predict how the current road scenario will change in the future. We successfully learn compact representations that use as few as 16 neural units for each of the two basic driving concepts we consider: cars and lanes. We maintain the two concepts separated in the latent space to facilitate the interpretation and manipulation of the perceptual representations. The source code for this paper is available at https://github.com/3lis/rnn_vae.

Driving requires the ability to handle unpredictable situations. Since it is not always possible to predict an impending danger, a good driver should preventively assess whether a situation has risks and adopt a safe behavior. Considering, in particular, the possibility of a pedestrian suddenly crossing the road, a prudent driver should limit the traveling speed. We present a work exploiting reinforcement learning to learn a function that specifies the safe speed limit for a given artificial driver agent. The safe speed function acts as a behavioral directive for the agent, thus extending its cognitive abilities. We consider scenarios where the vehicle interacts with a distracted pedestrian that might cross the road in hard-to-predict ways and propose a neural network mapping the pedestrian’s context onto the appropriate traveling speed so that the autonomous vehicle can successfully perform emergency braking maneuvers. We discuss the advantages of developing a specialized neural network extension on top of an already functioning autonomous driving system, removing the burden of learning to drive from scratch while focusing on learning safe behavior at a high-level. We demonstrate how the safe speed function can be learned in simulation and then transferred into a real vehicle. We include a statistical analysis of the network’s improvements compared to the original autonomous driving system. The code implementing the presented network is available at https://github.com/tonegas/safe-speed-neural-network with MIT license and at https://zenodo.org/communities/dreams4cars. [ABSTRACT FROM AUTHOR]

This paper deals with (deep) neural network implementations of sensorimotor control for automated driving. We show how to construct complex behaviors by re-using elementary neural network building blocks that can be trained and tested extensively; one of our goals is to mitigate the “black box” and verifiability issues that affect end-to-end trained networks. By structuring complex behaviors within a subsumption architecture, we retain the ability to learn (mostly at motor primitives level) with the ability to create complex behaviors by subsuming the (well-known) learned elementary perception-action loops. The learning process itself is simplified, since the agent needs only to learn elementary behaviors. At the same time, the structure imposed with the subsumption architecture ensures that the agent behaves in predictable ways (e.g., treating all obstacles uniformly). We demonstrate these ideas for longitudinal obstacle avoidance behavior, but the proposed approach can also be adapted to other situations. [ABSTRACT FROM AUTHOR]