Self-driving Research in Review: ICRA 2019 Digest
By: Peter Ondruska, Director, Research; and Guido Zuidhof and Lukas Platinsky, Software Engineers
Those of you who read our CVPR Digest know that we’ve begun sharing our internal team conference digests to make it easy to find self-driving resources from popular conferences.
This week, we’re sharing what we gathered from the International Conference on Robotics and Automation (ICRA) 2019. ICRA is one of the largest robotics conferences in the world. Out of the 1300 papers presented, we found over 250 directly or indirectly relevant to our self-driving efforts (4x more than CVPR — see the digest here). Read on for a summary of our findings.
Perception & Tracking
Distant Vehicle Detection Using Radar and Vision [video]
Paper from Oxford Robotics Institute — Detecting objects at a distance is difficult, but radar can help. This is one of the few papers that discusses radar-camera fusion to address the lack of radar data in public datasets.
SuperDepth: Self-Supervised, Super-Resolved Monocular Depth Estimation [video]
Paper from Toyota Research Institute — This paper discusses the state-of-the-art for achieving high-resolution monocular depth prediction.
I Can See Clearly Now: Image Restoration via De-Raining
Paper from Oxford Robotics Institute — Difficult weather conditions can create challenges for a self-driving perception system. This can be addressed by a Generative Adversarial Network (GAN) trained to remove image defects such as rain droplets on the camera.
Probably Unknown: Deep Inverse Sensor Modeling In Radar [video]
Paper from Oxford Robotics Institute — This paper explores training a neural network to decode radar readings into 2D occupancy maps using lidar supervision.
Prediction & Planning
Learning to Drive in a Day [video]
Paper from Wayve — Reinforcement learning is an approach that can, in theory, solve any problem through trial-and-error, but its application in robotics is difficult. This paper shows how to train an autonomous vehicle to drive in the real world using disengagement as a training signal. The car was able to learn to follow the road without any further supervisory training.
Motion Planning Networks: Bridging the Gap Between Learning-based and Classical Motion Planners [code] [website]
Paper from University of California, San Diego — The classical planning method of using hand-crafter algorithms might be slow for complex tasks. This can be fixed by neural networks that significantly speed them up while guaranteeing result quality.
Multimodal Trajectory Predictions for Autonomous Driving Using Deep Convolutional Networks
Paper from Uber ATG — Predicting what will happen is a key problem for autonomous driving. The authors trained an end-to-end neural network to learn to predict likely future trajectories of other traffic participants. Unlike previous works, it can predict several likely outcomes at once.
Learning from Demonstration in the Wild
Paper from Latent Logic — Learning from experts is a promising way for a self-driving car to learn to drive as opposed to hand-crafting such behaviors. The presented approach uses video from static traffic cameras to learn how people drive, and trains a policy that mimics the observed behaviors.
Mapping & Localisation
Visual SLAM: Why Bundle Adjust?
Paper from University of Adelaide and Queensland University of Technology — Sliding-window bundle adjustment has become a go-to method for SLAM applications since the paper titled Visual SLAM: Why Filter? was published. This new paper introduces rotational averaging into the SLAM world, a method used in recent Structure from Motion approaches. It provides a way to simplify SLAM systems and make them more robust to tackle challenging motions.
Learning Scene Geometry for Visual Localization in Challenging Conditions
Paper from Université Bourgogne Franche-Comte and Univ. Paris-Est — This paper explores a new way to utilize auxiliary depth information at training to make image descriptors represent geometric appearance. They compare several different methods and networks to obtain state-of-the-art results.
Lidar Measurement Bias Estimation via Return Waveform Modeling in a Context of 3D Mapping
Paper from Institut Pascal and Universite Laval — This paper outlines the impact of incidence angles on lidar measurements and provides a precise way to model it. The biases they compute can cause non-trivial mapping drift in some situations.
Simulation, Safety & Long-tail of Rare Events
Learning to Drive from Simulation without Real World Labels [video]
Paper from Wayve — Training a self-driving car in simulation as opposed to real-world is cheaper, faster and safer; however, such systems usually fail to generalize to the real world. The presented method shows how to learn to drive in simulation while using Cycle GAN to relate the real world with a simulated one. The system is tested on a car that can perform a simple behavior in the real world (following a lane) without any real-world training.
Generating Adversarial Driving Scenarios in High-Fidelity Simulators [video]
Paper from McGill University — Verifying performance of the driving algorithm is paramount for building safe self-driving cars. Authors describe method for automatic search of driving scenarios that can break a planning algorithm, and then use these scenarios to make the algorithm better.
Robustness to Out-of-Distribution Inputs via Task-Aware Generative Uncertainty
Paper from University of California, Berkeley — Neural networks can perform poorly when exposed to inputs that are different to the ones used for training. This can be addressed by both learning what is important in the input, and whether it was present in the training dataset.
[→ Read 10 more Simulation, Safety & Long-tail of Rare Events papers from ICRA]
Efficiency, Accuracy & Optimization
Where Should We Place LiDARs on the Autonomous Vehicle? — an Optimal Design Approach
Paper from Carnegie Mellon University — This paper explores a framework that models sensor configurations on a car with the purpose of finding the optimal placement.
Point Cloud Compression for 3D LiDAR Sensor using Recurrent Neural Network with Residual Blocks
Paper from Nagoya University — Storing raw lidar data from an autonomous vehicle can be expensive due to its size and volume. The paper proposes a trainable compression method that reduces the storage 100x and evaluates its impact on SLAM performance.
Large-Scale Object Mining for Object Discovery from Unlabeled Video
Paper from Aachen University — Discovering rare and useful information from data can be like searching for a needle in a haystack. This method learns to automatically detect and cluster new kinds of encountered objects in an unsupervised way from a large driving video collection.
[→ Read 2 more Efficiency, Accuracy & Optimization Papers from ICRA]
Ridesharing
Optimizing Vehicle Distributions and Fleet Sizes for Mobility-on-Demand
Paper from MIT — This paper explores optimizing the trade-offs of taxi fleet allocation to meet the rider demand in New York. Experiments suggest one needs 2864 vehicles to serve all the rides in the city.
Released Datasets
The H3D Dataset for Full-Surround 3D Multi-Object Detection and Tracking in Crowded Urban Scenes
Dataset from Honda Research Institute — This dataset contains data from camera, lidar, and 3D annotations of 160 highly crowded scenes around a vehicle containing 1M labelled instances in 27K images.
PedX: Benchmark Dataset for Metric 3D Pose Estimation of Pedestrians in Complex Urban Intersections
Dataset from Ford Center for Autonomous Vehicles Laboratory — This dataset contains 5K camera-lidar pairs with annotated pedestrian segmentation alongside full body-joint reconstructions, supplying accurate pedestrian poses for each annotation.
If this post caught your interest, we’re hiring! Check out our open roles here if you want to take on a part of the self-driving challenge, and be sure to follow our blog for more technical content.
