Robot Navigation Policy Learning with Automated Reward Shaping

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current robot navigation methods require tuning for each new robot and environment, and reinforcement learning can fail to produce robust policies due to sparse rewards, leading to inaccurate and non-robust navigation behaviors.

Innovation Solution

Implementing an automated reward shaping and network shaping technique using reinforcement learning to determine optimal reward weights and network architectures for policy networks, which enables the training of robust robotic navigation policies that can handle dynamic environments and obstacles effectively.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Extent of automation

If reinforcement learning is used with sparse rewards based on true objective, then the robot can learn navigation policies, but the policies are inaccurate and non-robust

Engineering Contradiction:
Improveautonomous navigation capabilityVSAvoidnavigation robustness
Core Design Contradiction:
Extent of automationVSReliability

Solution Approach 1:

The patent introduces an intermediary reward function that bridges the gap between sparse true objective rewards and dense guidance rewards. This intermediary reward provides intermediate feedback signals that guide the robot toward the goal without directly revealing the final objective, enabling more robust policy learning while maintaining automation.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent dynamically adjusts reward parameters including the intermediary reward weight, guidance reward weight, and exploration bonus weight during training. These parameter changes allow the system to balance exploration and exploitation, improving navigation robustness while maintaining autonomous capability.

Inventive Principle:
Principle #35Parameter changes

2Reliability

If hand-chosen reward shaping is used, then sparse rewards are addressed, but agents learn to exploit the reward function or terminate episodes early

Engineering Contradiction:
Improvereward signal effectivenessVSAvoidreward function design complexity
Core Design Contradiction:
ReliabilityVSEase of operation

Solution Approach 1:

The patent implements self-service through automated reward weight tuning where the system automatically adjusts intermediary reward weights and guidance reward weights based on training progress and performance metrics, eliminating the need for manual reward function design while preventing reward exploitation.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent incorporates feedback mechanisms that monitor training performance and automatically adjust reward parameters. The system uses feedback from navigation success rates, reward exploitation detection, and episode termination patterns to dynamically tune reward weights, preventing agents from learning to exploit the reward function.

Inventive Principle:
Principle #23Feedback

3Device complexity

If vision-based navigation with high-level motion primitives is used, then robot dynamics are abstracted away, but navigation requires tuning for each new robot and environment

Engineering Contradiction:
Improvenavigation system complexityVSAvoidenvironment adaptability
Core Design Contradiction:
Device complexityVSAdaptability or versatility

Solution Approach 1:

The patent creates a universal navigation policy through reinforcement learning that works across different robots and environments. The learned policy generalizes to various scenarios including different robot dynamics, environment types, and obstacle configurations, eliminating the need for separate tuning while maintaining low system complexity.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Solution Approach 2:

The patent implements dynamic reward parameter adjustment during training and deployment. The system adapts intermediary reward weights and exploration bonuses based on real-time performance feedback, enabling the navigation system to adjust to different robots and environments dynamically without requiring manual reconfiguration.

Inventive Principle:
Principle #15Dynamics

4Reliability

If motion planning that handles robot dynamics is used, then accurate navigation is achieved, but near-perfect sensor observations are required

Engineering Contradiction:
Improvenavigation accuracyVSAvoidsensor observation quality
Core Design Contradiction:
ReliabilityVSDifficulty of detecting and measuring

Solution Approach 1:

The patent introduces an intermediary reward function that provides robust guidance even with noisy sensor observations. This intermediary reward acts as a mediator between imperfect sensor data and navigation decisions, enabling accurate navigation without requiring near-perfect observations by providing stable intermediate feedback signals.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3837641B1Deep reinforcement learning-based techniques for end to end robot navigation
Publication Date: 2024.11.06 GOOGLE LLC
  • EP3837641B1 patent drawingFigure 1
  • EP3837641B1 patent drawingFigure 2A
  • EP3837641B1 patent drawingFigure 2B

AI summary

Using reinforcement learning to train a policy network that can be utilized, for example, by a robot in performing robot navigation and/or other robotic tasks. Various implementations relate to techniques for automatically learning a reward function for training of a policy network through reinforcement learning, and automatically learning a neural network architecture for the policy network.