Conditional variational auto encoder based dynamic motion for multitask imitation learning.

Area of Science:

• Robotics
• Machine Learning
• Control Systems

Background:

• Dynamic Motion Primitive (DMP) methods are effective for learning from demonstrations but often limited to single tasks.
• Existing deep learning frameworks for multi-task learning require extensive data and exhibit limited generalization to untrained states.

Purpose of the Study:

• To develop a novel framework integrating DMP advantages with conditional variational auto-encoders (cVAE) for enhanced multi-task learning.
• To enable robots to adapt learned behaviors to new goal positions and via-point constraints with minimal demonstrations.

Main Methods:

• A hybrid encoder-decoder architecture combining a dynamic system and a deep neural network.
• Deep neural networks generate task-conditioned torque, driving a dynamic system to produce desired trajectories.
• A fine-tuning method is proposed to ensure via-point constraints are met.

Main Results:

• The model demonstrates successful learning and adaptation on handwritten digit datasets and robotic manipulation tasks (pushing, reaching, grasping).
• Validation on a UR10 manipulator confirms effectiveness in real-world scenarios.
• Achieved 100% success in reaching and 93.33% in pushing/grasping tasks with only one demonstration per task.

Conclusions:

• The proposed DMP-cVAE framework offers a data-efficient and generalizable approach to robotic learning.
• It overcomes limitations of traditional DMP and deep learning methods, particularly in multi-task scenarios with limited data.
• The method shows significant promise for real-world robotic applications requiring adaptive motion generation.