Soft Actor-Critic Robot Control for Hybrid Action Exploration

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Reinforcement learning for non-prehensile object manipulation in hybrid action spaces, which include both discrete and continuous actions, is challenging due to limited exploration strategies in existing methods.

Innovation Solution

A soft actor-critic robot control model that incorporates entropy of both location and motion policies, using a diffusion model to enhance exploration, particularly in hybrid action spaces, by adapting loss values based on entropy and variance to improve learning diversity and generalization.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If existing reinforcement learning methods are used for non-prehensile manipulation in hybrid action spaces, then the learning process can handle both discrete and continuous actions, but the exploration capability is limited and insufficient for effective learning

Engineering Contradiction:
Improveexploration capabilityVSAvoidlearning process complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent applies parameter changes by modifying the loss function to include entropy terms for both location and motion policies. Specifically, the actor loss is formulated as L_actor = -E[Q(s, a)] - α_loc·H(loc_policy) - α_motion·H(motion_policy), where the entropy terms H(·) encourage exploration by penalizing deterministic policies. This parameter modification to the loss function directly enhances exploration capability while maintaining the hybrid action space framework.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent implements dynamics by making the exploration behavior adaptive through the diffusion model. The diffusion process dynamically adjusts the exploration trajectory by progressively denoising actions, allowing the system to transition from high exploration (early diffusion steps) to exploitation (later diffusion steps). This dynamic adjustment of exploration intensity resolves the contradiction by enabling effective exploration without requiring uniformly high complexity throughout the learning process.

Inventive Principle:
Principle #15Dynamics

2Adaptability or versatility

If entropy of location policy is considered during reinforcement learning, then exploration of discrete actions is improved, but the overall learning efficiency may be reduced due to additional computational requirements

Engineering Contradiction:
Improveexploration of discrete actionsVSAvoidlearning time
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent applies preliminary action by pre-computing the entropy of the location policy as part of the loss function formulation. The entropy H(loc_policy) = -Σ π(loc|s)·log(π(loc|s)) is calculated alongside the Q-value estimation, allowing the exploration incentive to be integrated into each training step without requiring separate computational passes. This preliminary incorporation of entropy calculation into the main learning loop minimizes additional time overhead.

Inventive Principle:
Principle #10Preliminary action

3Adaptability or versatility

If diffusion model is used to determine motion parameters as denoising sequence, then learning diversity is improved and generalization to out-of-distribution scenarios is enhanced, but the computational complexity increases

Engineering Contradiction:
Improvelearning diversity and generalizationVSAvoidcomputational complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent applies segmentation by decomposing the motion parameter determination into a sequence of denoising steps. Instead of directly outputting the final motion parameters, the diffusion model progressively refines noisy actions through multiple intermediate steps, where each step reduces the noise level. This segmentation of the action generation process into discrete denoising stages enables diverse exploration while maintaining computational tractability through incremental refinement.

Inventive Principle:
Principle #1Segmentation

4Adaptability or versatility

If both entropy of location policy and entropy of motion parameter policy are incorporated into loss calculation, then exploration for both discrete and continuous actions is enhanced, but the loss calculation and policy adaptation become more complex

Engineering Contradiction:
Improveexploration for discrete and continuous actionsVSAvoidloss calculation complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent applies merging by combining the entropy regularization terms for location and motion policies into a unified loss function. The total actor loss L_actor = -E[Q(s, a)] - α_loc·H(loc_policy) - α_motion·H(motion_policy) integrates both entropy penalties alongside the expected Q-value, allowing simultaneous optimization of exploration for both discrete location selection and continuous motion parameters. This merged formulation simplifies the overall optimization process compared to separate learning schemes.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP4706901A1Method of reinforcement learning of a soft actor-critic robot control model for non-prehensile manipulation of an object
Publication Date: 2026.03.11 ROBERT BOSCH GMBH
  • EP4706901A1 patent drawingFigure 1
  • EP4706901A1 patent drawingFigure 2
  • EP4706901A1 patent drawingFigure 3

AI summary

Various aspects relate to a method (200) of reinforcement learning a soft actor-critic robot control model (302) for manipulation of an object, wherein an actor (320) of the soft actor-critic robot control model (302) includes a motion parameter policy for determining, for each point of an object-representing point cloud (306), a motion parameter indicating how to move the object when being contacted at the point and includes a location policy for determining which point of the object-representing point cloud (306) is to be contacted, wherein the method includes a training of the soft actor-critic robot control model (302) using both, an entropy of the motion parameter policy and an entropy of the location policy at each point of the object-representing point cloud (306).