ChauffeurNet: Learning to Drive by Imitating the Best and Synthesizing the Worst

Abstract

Our goal is to train a policy for autonomous driving via imitation learning that is robust enough to drive a real vehicle. We find that standard behavior cloning is insufficient for handling complex driving scenarios, even when we leverage a perception system for preprocessing the input and a controller for executing the output on the car: 30 million examples are still not enough. We propose exposing the learner to synthesized data in the form of perturbations to the expert's driving, which creates interesting situations such as collisions and/or going off the road. Rather than purely imitating all data, we augment the imitation loss with additional losses that penalize undesirable events and encourage progress -the perturbations then provide an important signal for these losses and lead to robustness of the learned model. We show that the ChauffeurNet model can handle complex situations in simulation, and present ablation experiments that emphasize the importance of each of our proposed changes and show that the model is responding to the appropriate causal factors. Finally, we demonstrate the model driving a car in the real world.

SVG-Net: An SVG-based Trajectory Prediction Model

“…In [45], [39], [44], a Convolutional Neural Network (CNN) is used to extract scene's features. The learning power of CNN is utilized in [42], [23], [11], [7] to implicitly learn both interactions and the scene semantics. To this end, they render the scene semantics and states of agents in the scene in a multidimensional image and use CNNs to capture the underlying relations between dimensions.…”

“…By taking the input domain structure of neural networks in mind, it is clear that the scene's contextual information should be represented in a way that is digestible for the networks. In most of the previous works on vehicle trajectory prediction [39], [45], [23], [11], [7], scene's contextual information is rendered into image-like raster inputs, and 2D Convolutional Neural Networks (CNN) are employed to learn an abstract representation. This is inspired by the success of CNNs in various computer vision tasks [26], [21], [19], [41], making rendered images and CNNs as standard input representations and processors, respectively.…”

Self-Supervised Action-Space Prediction for Automated Driving

“…In relation to this distinction, a categorization by different input modalities to raster-, polyline-, or sensor-based approaches can be made. Raster-based approaches [10], [5], [11], [12], [13], [14] take in Bird's Eye View (BEV) images of an agent's environment with elements such as road lanes and bounding boxes of other perceived agents. By doing so, the CNN network can extract local spatial information from a single representational domain.…”

Structure Information is the Key: Self-Attention RoI Feature Extractor in 3D Object Detection

“…3D perception is an important part of the automatic driving system [1,11]. At present, many 3D detectors have been proposed.…”

“…The main contributions of this work can be summarized as (1) We indicate the essential difference between point-based and voxel-based RoI feature extractors is the tradeoff between location and structure information. Then we point out that compared to accurate position information, structural information is more important for 3D object detection; (2) We propose a plug-and-play module called self-attention RoI Feature Extractor (SARFE) to automatically adjust local features in a proposal to get features with stronger structure information; (3) Our proposed method SARFE achieves newly the-state-of-art performance on cyclist of KITTI dataset [8] while keeping real-time ability.…”