02_AI_Based.md

References

BASICS

1. LOAMS (LiDAR odometry & mapping)
   1. LOAM 2014: Lidar Odometry and Mapping in Real-time
   2. LOAM 2016: Low-drift and real-time lidar odometry and mapping
   3. V-LOAM 2015: Visual-lidar Odometry and Mapping: Low-drift, Robust, and Fast
   4. LeGO-LOAM 2018: Lightweight and Ground-Optimized Lidar Odometry and Mapping on Variable Terrain
      • Simplified LOAM (for ground vehicle) + Loop Closing
2. LIO (lidar- IMU odometry)
   1. LIO 2019: Tightly Coupled 3D Lidar Inertial Odometry and Mapping
   2. LIO-SAM 2020: Tightly-coupled Lidar Inertial Odometry via Smoothing and Mapping
      • LeGO-LOAM + IMU + GPS
   3. LVI-SAM 2021: Tightly-coupled Lidar-Visual-Inertial Odometry via Smoothing and Mapping
      • LIO-SAM + Vision
   4. DeepLIO 2021: Deep LIDAR Inertial Sensor Fusion for Odometry Estimation

ICRA 2021

1. Unsupervised Learning of Lidar Features for Use in a Probabilistic Trajectory Estimator

   • ICRA 2021 Best Student Paper Award.
   • Tim Barfoot - Autonomous Space Robotics Lab, UoT
   • Unsupervised Learning of Lidar Features

   

   
   * Detector Score: The higher the score, the more accurate the model is in its detections.

2. Radar Odometry Combining Probabilistic Estimation and Unsupervised Feature Learning

   • Same lab
   • Not ICRA, arxiv
     
     

3. Self-supervised Learning of LiDAR Odometry for Robotic Applications

   • Robotic Systems Lab, ETH Zurich
   • Geometric Loss
   
   

4. Interval-Based Visual-LiDAR Sensor Fusion

5. Dynamic Object Aware LiDAR SLAM based on Automatic Generationof Training Data

6. Self-Supervised Learning of Lidar Segmentation for Autonomous Indoor Navigation

UWB UoT, Prof. Angela Schoellig, Dynamic Systems Lab

2021: Learning-based Bias Correction for Time Difference of Arrival Ultra-wideband Localization of Resource-constrained Mobile Robots

Theory

• Due to the model nonlinearity, we use an M-estimation based extended Kalman filter (EKF) to estimate the states. Replacing the quadratic cost function in a conventional Kalman filter with a robust cost function ρ(·)—e.g. Geman- McClure (G-M), Huber or Cauchy.
  • The M estimators: try to reduce the effect of outliers by replacing the squared residuals r^2 by another function of the residuals, yielding: minⵉρ(r)
  • The standard least squares metho d tries to minimize minⵉ(r^2) which is unstable if there are outliers present in the data.

2020: Learning-based Bias Correction for Ultra-wideband Localizationof Resource-constrained Mobile Robots

Cited:

• Benchmark Dataset of Ultra-Wideband Radio Based UAV Positioning
• Automated Tuning of End-to-end Neural Flight Controllers for Autonomous Nano-drones
• Fully Onboard AI-powered Human-Drone Pose Estimation on Ultra-low Power Autonomous Flying Nano-UAVs
• Heading Estimation Using Ultra-wideband Received Signal Strength and Gaussian Processes
• Sensor Information Sharing Using a Producer-Consumer Algorithm on Small Vehicles

Theory

Two-step measurement correction:

1. Statistical outlier rejection

   • Using robot’s dynamics to filter inconsistent UWB range measurements: the maximum distance max a quadcopter can cover during time ∆t: dmax = ‖v∆t + 1 2 amax∆t2‖
   • statistical hypothesis: S = GPG + R (EKF: use the χ2 hypothesis test to determine whether a measurement innovation is likely coming from the distribution)

2. NN Bias compensation: spatially varying bias of TWR and TDoA measurements

   • Exclusively trained our NN with measurement whose actual bias less within a threshold of 0.7m

Notes

• Mitigation of UWB TWR measurement errors: (most of them leverage probabilistic methods)
  • Systematic biases
  • NLOS
  • Multipath propagation
• Since multi-path and NLOS propagation effects depend on a particular indoor environment, we only use the NN to explicitly model the pose-dependent bias.

Visual Localization UoT, Prof. Jonathan Kelly, STARS Lab

• 2020: Self-Supervised Deep Pose Corrections for Robust Visual Odometry
• 2018: VDPC-Net: Deep Pose Correction for Visual Localization

2018	2020

2020

They replace the supervised loss of DPC-Net with a photometric reconstruction loss that does not require any external ground truth pose information. Note: Self supervised vs unsupervised: There is a supervised training signal in these methods.

Tcorr corrects a classical VO estimate:
To parameterize this correction:

2018

Theory:

• In contrast to other methods that completely replace a classical visual estimator with a deep network, we propose an approach that uses a convolutional neural network to learn difficult-to-model corrections to the estimator from ground-truth training data.
• A novel loss function for learning SE(3) corrections based on a matrix Lie groups approach, with a natural formulation for balancing translation and rotation errors
• They use this loss to train a Deep Pose Correction network (DPC-Net) that predicts corrections for a particular estimator, sensor and environment.
• Others for loss function : based on SE(3) geodesic distance.
• Their loss naturally balances translation and rotation error without requiring a hand-tuned scalar hyper-paramete

cited:

Learning error models for graph SLAM

[Stereo Visual Odometry Pose Correction through Unsupervised Deep Learning] (https://starslab.ca/)

Cited

1- Learning Wheel Odometry and IMU Errors for Localization
2- A Smooth Representation of Belief over SO(3) for Deep Rotation Learning with Uncertainty
3- The Complex-Step Derivative Approximation on Matrix Lie Groups
4- Probabilistic Regression of Rotations using Quaternion Averaging and a Deep Multi-Headed Network
5- mproving Learning-based Ego-motion Estimation with Homomorphism-based Losses and Drift Correction
5- A Stacked LSTM-Based Approach for Reducing Semantic Pose Estimation ErrorRana

Comparision between different methods

• What s in My LiDAR Odometry Toolbox?
  

An Overview of Perception and Decision-Making in Autonomous Systems in the Era of Learning

• DeepLIO