Robotic Action Policy Transfer Across Simulation and Real-World Inputs

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Deep reinforcement learning algorithms are too slow to achieve performance on real-world robots and require significant computing resources, and existing approaches struggle to transfer learned policies from simulation to reality effectively, especially when dealing with changing network morphology or new input modalities.

Innovation Solution

A neural network system that includes a simulation-trained DNN with sufficient capacity and depth, and one or more robot-trained DNNs with minimal capacity, allowing for flexible task adaptation and efficient action selection by bridging the reality gap through progressive networks, enabling faster learning and reduced computational footprint.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If deep reinforcement learning algorithms are used to train policies on real-world robots, then the agent can learn complex tasks, but the training process is too slow and requires significant computing resources

Engineering Contradiction:
Improvelearning speedVSAvoidcomputing resources
Core Design Contradiction:
ProductivityVSUse of energy by moving object

Solution Approach 1:

The training process is segmented into two distinct phases: simulation training and real-robot training. The simulation-trained DNN handles the bulk of complex learning in a virtual environment, while the robot-trained DNN with minimal capacity performs only the final adaptation on the real robot. This segmentation allows most computationally intensive work to occur in simulation where resources are abundant, reducing the burden on real-robot computing resources.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

A simulation environment acts as an intermediary between theoretical reinforcement learning algorithms and real-world robot execution. The simulation-trained DNN serves as a bridge, pre-processing complex learning tasks in a controlled virtual environment before transferring the learned policies to the actual robot, thereby accelerating real-world training while reducing computational requirements on the physical system.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Adaptability or versatility

If a simulation-trained DNN with sufficient capacity and depth is used, then the network can learn tasks from scratch in simulation, but transferring to real robots requires accommodating dissimilar inputs and changing network morphology

Engineering Contradiction:
Improvetask learning capabilityVSAvoidnetwork structure complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The system dynamically adapts network capacity across different training phases. The simulation-trained DNN operates with full capacity and depth to learn complex tasks from scratch. When transferring to the real robot, the system switches to a robot-trained DNN with minimal capacity that is dynamically adjusted to match the specific requirements of the physical environment, allowing flexible accommodation of dissimilar inputs without maintaining fixed complex architecture.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

Different parts of the neural network system have different capacities optimized for their specific functions. The simulation-trained DNN has sufficient capacity and depth for comprehensive task learning in simulation, while the robot-trained DNN has minimal capacity tailored specifically for real-robot adaptation. This local optimization allows each component to be precisely matched to its operational requirements rather than using a uniform architecture throughout.

Inventive Principle:
Principle #3Local quality

3Adaptability or versatility

If existing approaches are used for policy transfer, then training can occur on real robots, but they are unable to accommodate changing network morphology or new input modalities

Engineering Contradiction:
Improveinput modality flexibilityVSAvoidtraining time
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The system performs preliminary training in simulation where the agent learns tasks from scratch using a fully-capable DNN. This preliminary action in the virtual environment prepares the policy before real-robot training begins, allowing the actual robot training to focus only on adapting to physical realities rather than learning basic task concepts, thereby accommodating new input modalities without losing time on fundamental learning.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3523758B1Neural networks for selecting actions to be performed by a robotic agent
Publication Date: 2024.10.02 DEEPMIND TECH LTD
  • EP3523758B1 patent drawingFigure 1
  • EP3523758B1 patent drawingFigure 2
  • EP3523758B1 patent drawingFigure 3

AI summary

A system includes a neural network system implemented by one or more computers. The neural network system is configured to receive an observation characterizing a current state of a real-world environment being interacted with by a robotic agent to perform a robotic task and to process the observation to generate a policy output that defines an action to be performed by the robotic agent in response to the observation. The neural network system includes: (i) a sequence of deep neural networks (DNNs), in which the sequence of DNNs includes a simulation-trained DNN that has been trained on interactions of a simulated version of the robotic agent with a simulated version of the real-world environment to perform a simulated version of the robotic task, and (ii) a first robot-trained DNN that is configured to receive the observation and to process the observation to generate the policy output.