Autonomous Driving Policy Learning With Guided Vision-Based Search

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Rule-based methods in autonomous driving are not scalable, and reinforcement learning methods based on Q-learning or policy gradient are not sample efficient, leading to time-consuming solutions due to the need to explore a large state space.

Innovation Solution

A vision-based reinforcement learning framework that combines environmental information from cameras with guided policy search, using deep learning neural networks to iteratively optimize local policies for specific driving tasks and train a general global policy, limiting the search space to be sample efficient.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If reinforcement learning methods based on Q-learning or policy gradient are used for autonomous driving, then the system can learn optimal driving policies, but the training time becomes very long due to the need to explore a large state space

Engineering Contradiction:
Improvelearning effectivenessVSAvoidtraining time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent applies preliminary action by using imitation learning to pre-train the policy network with demonstration data before performing reinforcement learning. This preliminary training phase allows the system to learn from expert demonstrations without extensive exploration, significantly reducing the subsequent RL training time while maintaining learning effectiveness.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent segments the training process into two distinct phases: imitation learning phase and reinforcement learning phase. The imitation learning phase uses demonstration data to initialize the policy, while the RL phase fine-tunes it. This segmentation allows each phase to focus on specific learning objectives, improving overall efficiency.

Inventive Principle:
Principle #1Segmentation

2Adaptability or versatility

If extensive exploration of the state space is performed to ensure comprehensive learning, then the learning coverage is improved, but the training time and computational resources increase significantly

Engineering Contradiction:
Improvelearning coverageVSAvoidtraining efficiency
Core Design Contradiction:
Adaptability or versatilityVSProductivity

Solution Approach 1:

By performing preliminary imitation learning with diverse demonstration data, the system achieves broad learning coverage before RL training. The demonstration data covers various driving scenarios and states, allowing the policy to be initialized with comprehensive knowledge without requiring extensive RL exploration.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent introduces demonstration data as an intermediary between the agent and the environment. Instead of directly exploring the state space through trial-and-error RL, the agent first learns from demonstrated trajectories, which serve as a bridge providing efficient learning coverage across diverse states.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Ease of operation

If rule-based methods are used for autonomous driving, then the system can operate with clear decision logic, but the system cannot scale well to handle diverse driving scenarios

Engineering Contradiction:
Improvedecision logic clarityVSAvoidscalability
Core Design Contradiction:
Ease of operationVSAdaptability or versatility

Solution Approach 1:

The patent replaces rule-based mechanical decision logic with a learned policy network. Instead of manually encoding driving rules, the system learns optimal driving behaviors through imitation learning and reinforcement learning from state observations and actions, enabling scalability to diverse scenarios while maintaining operational effectiveness.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent transitions from fixed rule-based parameters to learned continuous policies. The policy network outputs continuous control signals based on learned patterns from demonstration and RL training, allowing flexible adaptation to various driving scenarios without requiring explicit rule changes for each case.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3564861B1Vision-based sample-efficient reinforcement learning framework for autonomous driving
Publication Date: 2024.02.07 SONY GROUP CORP
  • EP3564861B1 patent drawingFigure 1~3
  • EP3564861B1 patent drawingFigure 4
  • EP3564861B1 patent drawingFigure 5

AI summary

A framework combines vision and sample-efficient reinforcement-learning based on guided policy search for autonomous driving. A controller extracts environmental information from vision and is trained to drive using reinforcement learning.