Hybrid Imitation Learning Framework for Robotic Manipulation Tasks

Area of Science:

• Robotics
• Machine Learning
• Artificial Intelligence

Background:

• Robotic manipulation tasks require efficient learning methods.
• Current imitation learning methods like behavior cloning (BC) and state cloning (SC) have limitations in training efficiency and policy flexibility.
• Hybrid approaches are being explored to overcome these limitations.

Purpose of the Study:

• To propose a novel hybrid imitation learning (HIL) framework combining BC and SC for enhanced robotic manipulation task learning.
• To improve training efficiency and policy flexibility in robotic learning.
• To demonstrate the effectiveness of the HIL framework through experimental validation.

Main Methods:

• A hybrid imitation learning (HIL) framework integrating BC and SC.
• Adaptive loss mixing to combine BC and SC losses.
• Pretrained dynamics networks to enhance SC efficiency.
• Stochastic state recovery for stable policy network learning.

Main Results:

• The HIL framework demonstrated approximately 2.6 times higher performance improvement than pure BC.
• The HIL framework achieved approximately four times faster training time than pure SC.
• Compared to BC + Reinforcement Learning (RL), HIL showed 1.6 times higher performance improvement and 2.2 times faster training.

Conclusions:

• The proposed HIL framework offers significant improvements in training efficiency and policy flexibility for robotic manipulation tasks.
• HIL provides a superior alternative to pure BC, pure SC, and BC + RL methods.
• This framework advances the field of imitation learning for complex robotic applications.