Robot Action Sequence Prediction for Adaptive Task Planning

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Robotic systems face challenges in performing complex object manipulation tasks due to the complexity of task planning over long horizons and adapting to changing environments, as existing technologies struggle to learn dependencies and transitions across subtasks effectively.

Innovation Solution

An environment-conditioned action sequence prediction model using a convolutional neural network and sequence-to-sequence model processes vision data to predict and adapt sequences of actions, allowing robots to interactively adjust to environmental changes and recover from failures by learning from demonstrations and visual observations.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If a robot uses traditional programming to perform tasks, then it can execute specific manipulations reliably, but it cannot adapt to changes in the environment such as objects in different positions or different initial states

Engineering Contradiction:
Improveadaptability to environmental changesVSAvoidcomplexity of task planning
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent segments complex robotic tasks into sequences of subtasks, where each subtask corresponds to a specific action (e.g., pick object, move object, place object). This segmentation allows the robot to handle complex tasks through manageable steps while maintaining adaptability to environmental variations.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system performs preliminary learning of task sequences from demonstrations before actual execution. By pre-learning the sequence of actions needed for various tasks from video demonstrations, the robot builds a knowledge base that enables adaptive execution without requiring complex real-time planning.

Inventive Principle:
Principle #10Preliminary action

2Adaptability or versatility

If a robot learns task sequences from demonstrations, then it can generalize to new situations, but it requires complex processing to learn dependencies and transitions across subtasks

Engineering Contradiction:
Improveability to generalize to new situationsVSAvoidcomplexity of learning dependencies
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The system uses feedback from video demonstrations to learn task sequences. By analyzing demonstrated tasks and extracting action sequences, the system automatically learns the dependencies and transitions between subtasks without explicit programming, reducing the complexity of learning while improving generalization.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent introduces an intermediate representation layer between raw video demonstrations and robot execution. This intermediate layer processes demonstrations into structured task sequences, making the learning process more efficient and the dependencies between subtasks more manageable.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Productivity

If a robot executes long-horizon plans, then it can complete complex multi-step tasks, but it struggles to maintain task completion when encountering unexpected environmental changes

Engineering Contradiction:
Improvetask completion capabilityVSAvoidresilience in dynamic situations
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The system dynamically adjusts task execution based on observed environmental changes. Rather than rigidly following pre-planned sequences, the robot can modify its action sequence in real-time based on what it observes, allowing it to complete long-horizon tasks while adapting to unexpected changes.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The robot performs self-correction by monitoring its own execution and the environment. When deviations occur from the planned sequence, the system automatically adjusts its behavior to recover and complete the task, reducing reliance on external intervention and improving reliability.

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP4003665B1Determining environment-conditioned action sequences for robotic tasks
Publication Date: 2024.11.13 GOOGLE LLC
  • EP4003665B1 patent drawingFigure 1
  • EP4003665B1 patent drawingFigure 2A~2B
  • EP4003665B1 patent drawingFigure 2C

AI summary

Training and/or using a machine learning model for performing robotic tasks is disclosed herein. In many implementations, an environment-conditioned action sequence prediction model is used to determine a set of actions as well as a corresponding particular order for the actions for the robot to perform to complete the task. In many implementations, each action in the set of actions has a corresponding action network used to control the robot in performing the action.