Robotic Imitation Learning with Task Sequencing for Long-Horizon Manipulation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing imitation learning methods in robotics struggle with complex and long-horizon object manipulation tasks, as they require explicit robot states and actions, and fail to accommodate skill learning from video demonstrations effectively.

Innovation Solution

A decoupled neural network architecture comprising a Task Sequencing Network (TSNet) and a Task Agnostic Dynamic Movement Primitive Library (TADL) is used to segment and execute primitive actions in robotics. TSNet segments task primitives from video demonstrations using a hybrid CNN-RNN architecture with CTC loss, while TADL executes these primitives sequentially on a robotic arm.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If explicit robot states and actions are used in imitation learning, then the learning process follows traditional LfD methodologies, but the system fails to accommodate skill learning from video demonstrations and struggles with long-horizon tasks

Engineering Contradiction:
Improveability to learn from video demonstrationsVSAvoidcomplexity of handling long-horizon tasks
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent segments long-horizon tasks into multiple short-horizon sub-tasks or primitives. Each primitive is learned independently from video demonstrations using pose estimation and temporal segmenting, then composed sequentially to execute complex tasks. This segmentation allows the system to handle long-horizon tasks by breaking them down into manageable units that can be learned from video without requiring explicit robot state annotations.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces an intermediary representation layer between video input and robot execution. This includes pose estimation of objects and human body parts, temporal segmenting to identify action boundaries, and primitive selection mechanisms. These intermediaries enable the system to learn from unannotated video demonstrations and translate them into robot-executable primitives, bridging the gap between visual observation and physical execution.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Ease of operation

If end-to-end approaches are used for visual imitation learning, then image space is directly mapped to robot actions, but this reduces efficiency, scope of task customization, and makes training harder

Engineering Contradiction:
Improveease of learning by scene observationVSAvoidtraining efficiency and task customization
Core Design Contradiction:
Ease of operationVSProductivity

Solution Approach 1:

The patent segments the end-to-end learning process into distinct modular components: video processing, pose estimation, temporal segmenting, primitive selection, and execution. Each module performs a specific function and can be independently trained or customized. This modular segmentation maintains the simplicity of learning from scene observation while improving training efficiency and task customization capability.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces dynamic primitive selection based on the current task context and observed human actions. Instead of a fixed mapping from images to actions, the system dynamically selects appropriate primitives from a library based on the segmented temporal structure and pose estimation results. This dynamic approach enables task customization while maintaining ease of operation through video observation.

Inventive Principle:
Principle #15Dynamics

3Manufacturing precision

If DMPs are used for short-horizon manipulation tasks, then successful trajectory reproduction is achieved, but their use in long-horizon tasks remains a challenge

Engineering Contradiction:
Improvetrajectory reproduction accuracyVSAvoidtask horizon duration
Core Design Contradiction:
Manufacturing precisionVSDuration of action of moving object

Solution Approach 1:

The patent segments long-horizon tasks into sequences of short-horizon primitives, each executable by DMPs. Each primitive represents a temporally bounded action (e.g., reach, grasp, move, place) that can be accurately executed using traditional DMP trajectory reproduction. The segmentation ensures that no single DMP needs to handle the entire long-horizon task,而是 each handles a manageable sub-task with high precision.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent ensures continuous execution of DMP primitives by sequentially chaining them together based on temporal segmenting results. Each primitive execution flows into the next without interruption, maintaining continuous useful action throughout the long-horizon task. This sequential composition allows DMPs to maintain their trajectory accuracy while handling extended task durations through continuous primitive execution.

Inventive Principle:
Principle #20Continuity of useful action

Data Source

PatentEP4549107A1System and method for imitation learning in robotics for complex task learning
Publication Date: 2025.05.07 TATA CONSULTANCY SERVICES LTD
  • EP4549107A1 patent drawingFigure 1
  • EP4549107A1 patent drawingFigure 2A
  • EP4549107A1 patent drawingFigure 2B

AI summary

The embodiments of present disclosure herein address unresolved problem of key capabilities that an intelligent robot must possess to perform complex and long-horizon object manipulation tasks without human intervention. The present disclosure overcomes above mentioned challenges by decoupling imitation learning task into two parts using a decouple neural network architecture comprising two modules. A high-level module that segments out sequence of primitive actions from video of human demonstration of a complex task and a low-level module that executes those primitives sequentially on robot. The high-level module, Task Sequencing Network (TSNet), is a hybrid neural network composed of Convolutional Neural Network (CNN), Recurrent Neural Network (RNN) with Connectionist Temporal Classification (CTC) loss, while low-level controller is built using dynamic movement primitives (DMP). The network is trained in a supervised learning setting. The CNN, RNN with CTC combination helps in efficiently segmenting task primitives in sequence for given input video demonstrations.