Robot Action Correction Learning From Human Feedback

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Robots often perform actions incorrectly due to inaccurate models and dynamic environments, and may not recognize these errors, leading to incorrect parameter determination and ineffective performance.

Innovation Solution

A method to generate correction instances based on human input, which include sensor data and incorrect parameter information, to train and refine neural network models used by robots, allowing for real-time adaptation and improvement of action performance across multiple robots in diverse environments.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If robots use pre-trained models to perform actions autonomously, then productivity is improved, but manufacturing precision deteriorates due to model inaccuracies and inability to recognize errors

Engineering Contradiction:
Improveautonomous action performanceVSAvoidaction accuracy
Core Design Contradiction:
ProductivityVSManufacturing precision

Solution Approach 1:

The system introduces human feedback through correction input that provides ground truth information about correct actions. This feedback loop allows the robot to learn from human corrections and improve action accuracy over time while maintaining autonomous operation. The correction instances containing human-provided correct parameters serve as feedback signals that refine the policy model.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The system performs preliminary training using correction instances collected from multiple robots before deploying updated models. By pre-training models with aggregated correction data from diverse sources, the system prepares improved action policies in advance, allowing robots to benefit from collective learning experiences without interrupting their autonomous productivity.

Inventive Principle:
Principle #10Preliminary action

2Adaptability or versatility

If robots operate in dynamic environments with varied conditions, then adaptability is improved, but reliability deteriorates due to model inaccuracies and unrecognized errors

Engineering Contradiction:
Improveenvironmental adaptabilityVSAvoidaction correctness
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The system creates universal correction instances that can be applied across multiple robots and diverse environments. By collecting and training on correction data from various environmental conditions and robot types, the system develops generalized models that maintain reliability across different contexts while preserving adaptability to specific situations.

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

Solution Approach 2:

The system merges correction instances from multiple robots operating in different environments to create a comprehensive training dataset. This combination of diverse correction data allows the model to learn robust patterns that improve reliability across varied conditions while maintaining the ability to adapt to specific environmental nuances.

Inventive Principle:
Principle #5Merging (Combining)

3Ease of operation

If robots rely on control processes to determine action correctness, then ease of operation is improved, but measurement precision deteriorates because control processes cannot detect actual errors

Engineering Contradiction:
Improveautonomous error detectionVSAvoiderror recognition accuracy
Core Design Contradiction:
Ease of operationVSMeasurement precision

Solution Approach 1:

The system introduces human users as intermediaries who provide correction input to bridge the gap between robot control processes and actual action correctness. Humans serve as mediators who can accurately assess whether actions are correct, compensating for the limitations of automated control processes while maintaining ease of autonomous operation.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3628031B1Determining and utilizing corrections to robot actions
Publication Date: 2023.06.07 X DEVELOPMENT LLC
  • EP3628031B1 patent drawingFigure 1
  • EP3628031B1 patent drawingFigure 2A
  • EP3628031B1 patent drawingFigure 2B

AI summary

Methods, apparatus, and computer-readable media for determining and utilizing human corrections to robot actions. In some implementations, in response to determining a human correction of a robot action, a correction instance is generated that includes sensor data, captured by one or more sensors of the robot, that is relevant to the corrected action. The correction instance can further include determined incorrect parameter(s) utilized in performing the robot action and/or correction information that is based on the human correction. The correction instance can be utilized to generate training example(s) for training one or model(s), such as neural network model(s), corresponding to those used in determining the incorrect parameter(s). In various implementations, the training is based on correction instances from multiple robots. After a revised version of a model is generated, the revised version can thereafter be utilized by one or more of the multiple robots.