Robot Control Policy Training With Bellman Uncertainty Variance

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Reinforcement learning methods face inefficiencies in exploration and sample-efficiency due to over-approximation of value function variance, leading to suboptimal control policy training, especially in scenarios with high epistemic uncertainty.

Innovation Solution

Solving the Bellman uncertainty equation with a revised reward function that sets the difference between total uncertainty about the mean of subsequent states and average aleatoric uncertainty, allowing for tighter variance estimates and biasing the control policy towards regions with higher variance, thereby improving exploration and stability.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If the Bellman uncertainty equation is solved with the conventional reward function (total uncertainty), then an upper bound of value function variance is obtained, but the estimation is loose and not tight

Engineering Contradiction:
Improvevariance estimation accuracyVSAvoidreward function complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent changes the parameter definition of the reward function from total uncertainty to the difference between total uncertainty about the mean and average aleatoric uncertainty. This parameter transformation enables tight variance estimation by isolating the epistemic uncertainty component that truly reflects model uncertainty, while filtering out irreducible aleatoric noise.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent segments the total uncertainty into two distinct components: total uncertainty about the mean value and average aleatoric uncertainty. By decomposing the uncertainty structure, the method can selectively utilize the epistemic uncertainty portion for exploration guidance while accounting for aleatoric uncertainty in the variance calculation, achieving both accuracy and interpretability.

Inventive Principle:
Principle #1Segmentation

2Productivity

If reinforcement learning uses standard exploration methods, then the policy may exploit known rewards, but exploration efficiency is low and sample-efficiency deteriorates

Engineering Contradiction:
Improvetraining efficiencyVSAvoidepisodes required for training
Core Design Contradiction:
ProductivityVSLoss of time

Solution Approach 1:

The patent implements a feedback mechanism where the estimated value function variance (derived from the revised Bellman uncertainty equation) continuously guides the exploration policy. This feedback loop enables the agent to dynamically adjust exploration behavior based on current uncertainty estimates, prioritizing regions with high epistemic uncertainty and thereby improving training efficiency and reducing the number of episodes required.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent performs preliminary uncertainty estimation and variance calculation before policy execution. By pre-computing the value function variance and identifying high-uncertainty regions in advance, the system can proactively guide exploration towards promising areas, avoiding wasted episodes on already-explored or low-value regions.

Inventive Principle:
Principle #10Preliminary action

3Adaptability or versatility

If the control policy is biased towards high uncertainty regions, then exploration improves, but the training complexity increases

Engineering Contradiction:
Improveexploration capabilityVSAvoidpolicy training complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent applies local quality by biasing the control policy specifically in high-uncertainty regions while maintaining standard behavior in low-uncertainty regions. The value function variance estimation enables spatially differentiated policy behavior, enhancing exploration capability where needed without unnecessarily complicating the policy in already-well-explored areas.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP4386632A1Device and method for controlling a robot
Publication Date: 2024.06.19 ROBERT BOSCH GMBH
  • EP4386632A1 patent drawingFigure 1
  • EP4386632A1 patent drawingFigure 2~3
  • EP4386632A1 patent drawing

AI summary

According to various embodiments, a method for training a control policy is described, comprising estimating the variance of a value function which associates a state with a value of the state or a pair of state and action with a value of the pair by solving a Bellman uncertainty equation, wherein, for each of multiple states, the reward function of the Bellman uncertainty equation is set to the difference of the total uncertainty about the mean of the value of the subsequent state following the state and the average aleatoric uncertainty of the value of the subsequent state and biasing the control policy in training towards regions for which the estimation gives a higher variance of the value function than for other regions.