Vision-Based Robot Motion Planning With Object-Centric Latent States

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing reinforcement learning (RL) methods struggle with long-horizon decision-making problems in robot motion planning, particularly in complex environments with sparse rewards, and face challenges in efficiently generating models from high-dimensional visual data.

Innovation Solution

The method employs object-centric latent states and a rapidly exploring random tree (RRT) approach combined with a diffusion probabilistic model for global sampling, using a graph neural network to enhance robot motion planning efficiency and robustness, particularly in vision-based robotic manipulation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If reinforcement learning is used for long-horizon decision-making in robot motion planning, then the robot can solve complex manipulation tasks, but the sample efficiency is poor and learning converges slowly due to sparse rewards

Engineering Contradiction:
Improvemotion planning capabilityVSAvoidlearning convergence time
Core Design Contradiction:
ProductivityVSLoss of time

Solution Approach 1:

The patent pre-trains a world model to predict future latent states before actual RL training. This preliminary action allows the system to learn environmental dynamics and object interactions in advance, providing a strong foundation that accelerates subsequent RL convergence and reduces sample efficiency issues during actual task learning.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent introduces a latent state representation as an intermediary between high-dimensional visual observations and RL decision-making. This latent space compresses visual information into meaningful object states, enabling the world model to efficiently predict future states and providing the RL agent with compact, informative state representations that improve learning efficiency.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Measurement precision

If high-dimensional visual data is used for robot perception, then the robot can accurately understand complex environments, but the computational complexity and data processing burden increase significantly

Engineering Contradiction:
Improveenvironment perception accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent extracts only the essential features from high-dimensional visual data by training a world model to predict latent states of objects. This extraction process identifies and retains only the most relevant environmental information needed for motion planning, discarding redundant visual details and significantly reducing computational complexity while maintaining perception accuracy.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent transforms visual data from high-dimensional pixel space into a compressed latent state space with fewer dimensions. This parameter transformation changes the representation from raw pixels to meaningful object states, reducing computational burden while preserving the essential information needed for accurate environment understanding and robot planning.

Inventive Principle:
Principle #35Parameter changes

3Productivity

If traditional motion planning algorithms are used, then the planning process is computationally efficient, but they struggle to find feasible paths over long temporal horizons in complex environments

Engineering Contradiction:
Improveplanning efficiencyVSAvoidfeasibility of long-horizon paths
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent employs a dynamic world model that can predict future latent states at multiple time steps into the future. This dynamic prediction capability allows the planner to evaluate long-horizon trajectories by simulating their outcomes, enabling reliable feasibility assessment of extended paths while maintaining computational efficiency through the compact latent state representations.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP4691702A1Method for vision-based robot motion planning using object-centric latent states
Publication Date: 2026.02.11 ROBERT BOSCH GMBH
  • EP4691702A1 patent drawingFigure 1
  • EP4691702A1 patent drawingFigure 2
  • EP4691702A1 patent drawing

AI summary

Various aspects relate to a method (200) for robot motion planning, the method including: receiving (202) an image sequence which ends at a point in time and shows objects in a surrounding of a robot (101) during a prior control of the robot; inputting (204) the image sequence into an encoder to predict a current latent state which represents a respective latent object state for each object of the objects at the point in time; generating (206), using a rapidly exploring random tree, RRT, approach, a latent state graph starting from the current latent state, wherein each predicted latent state of the latent state graph represents a respective predicted latent object state for each object of the objects; determining (208), by applying a policy for solving a Markow Decision Process on the latent state graph, a control trajectory for controlling the robot to achieve a goal state.