Ground Robot Visual Odometry Using Action Priors and Coarse Pose
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing visual odometry methods for robot navigation in noisy indoor environments suffer from sample inefficiency, embodiment specificity, and dataset specificity, requiring large compute resources and failing to perform well in wide-baseline settings with noisy RGB-D observations and actuations.
Innovation Solution
A method and system for motion prior based visual odometry using a Geometric Coarse Pose Estimator (GCPE) and Neural Fine Pose Regression (NFPR) modules to estimate a robot's relative pose, incorporating action priors to improve localization and navigation, with a lightweight CNN-based approach for efficient training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If learned VO methods are used with RL-based navigation policies, then robust performance is achieved, but sample efficiency deteriorates and huge compute resources are required
Solution Approach 1:
The patent segments the visual odometry problem into two distinct modules: a geometric coarse pose estimator that handles wide-baseline geometry, and a neural fine pose regressor that refines the estimate. This segmentation allows each module to be optimized independently, improving sample efficiency while maintaining robust performance.
Solution Approach 2:
The geometric coarse pose estimator performs preliminary pose estimation using only geometric constraints before the neural network refines it. This preliminary action provides a good initial estimate that reduces the learning burden on the neural network, improving sample efficiency.
2Reliability
If learned VO methods are used, then robust performance is achieved, but embodiment specificity and dataset specificity increase
Solution Approach 1:
The patent designs the neural fine pose regressor to work with generic RGB-D inputs without being tied to specific robot embodiments or datasets. The geometric module uses universal geometric constraints that apply across different embodiments, enabling zero-shot transfer to new robots and environments.
Solution Approach 2:
The patent changes the approach from learning embodiment-specific features to learning geometry-invariant parameters. The geometric coarse pose estimator uses parameters like baseline distance and viewing angles that are universal across different robot embodiments, improving generalizability.
3Measurement precision
If classical or learned VO methods are used, then localization is attempted, but they fail in wide-baseline settings with noisy RGB-D observations
Solution Approach 1:
The patent applies beforehand cushioning by using the geometric coarse pose estimator to provide a reliable initial pose estimate before the neural network refinement. This geometric prior cushions against the effects of noisy RGB-D observations, ensuring stable localization even in challenging conditions.
Solution Approach 2:
The geometric coarse pose estimator acts as an intermediary between the noisy RGB-D observations and the neural fine pose regressor. It mediates the information flow by providing a cleaned, geometrically-consistent initial estimate that reduces the impact of observations noise on the final localization accuracy.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Traditional Visual Odometry (VO) systems face challenges in wide-baseline scenarios leading to drift and catastrophic failures in robot navigation and suffer from sample inefficiency during training. Embodiments of present disclosure provide a method and system for motion prior based VO for ground robot navigation. A robot acquires a previous observation at a previous time step, an action prior, and a current observation at a current time step. Using this data, VO module estimates relative pose between previous and current observations. The VO module comprises a Geometric Coarse Pose Estimator module (GCPE) to estimate a coarse relative pose, and Neural Fine Pose Regression model (NFPR) to regress fine pose using the coarse pose and the acquired inputs. Then, using the relative pose, the current goal coordinates are updated. Finally, a navigation policy determines next action to take to reach the goal using updated goal coordinates and the current observation.