Camera Policy Neural Networks for Occlusion-Aware Active Vision

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing machine learning models struggle to effectively control the position of camera sensors on robots to optimize sensor readings without relying on external rewards or labels, particularly in noisy environments, leading to inefficiencies and occlusions.

Innovation Solution

Training a camera policy neural network to adjust the position of a camera sensor based on target sensor inputs and image observations, using reinforcement learning and sensor prediction neural networks to generate accurate predictions and rewards, allowing the network to learn active vision skills that avoid occlusions and frame sensors optimally.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If a camera policy neural network is trained using external rewards or labels, then training efficiency is improved, but the system becomes dependent on external reward functions that may not accurately reflect true sensor reading quality

Engineering Contradiction:

Improvetraining efficiencyVSAvoidreward accuracy

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The system trains the camera policy neural network using self-generated rewards based on sensor prediction errors. The sensor prediction neural network generates predictions from camera images, and the reward is computed from the error between these predictions and actual sensor readings. This self-service approach eliminates dependency on external reward functions while maintaining training efficiency through automated reward generation.

Inventive Principle:

Principle #25Self-service

2Measurement precision

If the camera sensor position is adjusted frequently to optimize sensor readings, then the quality of sensor data is improved, but the system becomes more vulnerable to occlusions and environmental noise

Engineering Contradiction:

Improvesensor reading qualityVSAvoidocclusion vulnerability

Core Design Contradiction:

Measurement precisionVSObject-affected harmful factors

Solution Approach 1:

The system implements a feedback loop where sensor prediction errors are used to compute rewards that guide the camera policy neural network. The camera adjusts its position based on predicted sensor readings and the resulting prediction errors, creating a closed-loop system that continuously optimizes viewing angles to avoid occlusions while maintaining measurement precision.

Inventive Principle:

Principle #23Feedback

3Adaptability or versatility

If a sensor prediction neural network is used to generate predictions from camera images, then the system can learn active vision skills, but the computational complexity increases

Engineering Contradiction:

Improveactive vision capabilityVSAvoidneural network complexity

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The system segments the vision system into two separate neural networks: a sensor prediction neural network that processes camera images to generate sensor readings, and a camera policy neural network that processes these predictions to determine camera position adjustments. This segmentation allows each network to be optimized independently, managing overall system complexity while enabling sophisticated active vision capabilities.

Inventive Principle:

Principle #1Segmentation

Data Source

PatentUS20250365502A1Training camera policy neural networks through self-prediction

Publication Date: 2025.11.27 GDM HOLDING LLC

AI summary

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training a camera policy neural network.