his video discusses a novel imitation learning framework designed for the joint training of teacher and student policies, addressing the challenge of partial observability faced by students. By incorporating an imitation upper bound, the framework optimizes the teacher policy while ensuring it can be imitated effectively by the student. The method shows promise across various tasks, including maze navigation, quadrotor obstacle avoidance, and robotic manipulation.
Imitation Learning Framework.
The framework enables joint training of teacher and student policies, aiming to improve the student’s ability to imitate complex behaviors despite limited observability.
Teacher-Student Asymmetry Challenge.
A key challenge in privileged imitation learning is that a student may not be able to fully imitate the teacher’s actions due to differences in observable information.
Imitation Upper Bound Addition.
The framework introduces an imitation upper bound that relates teacher and student performance, guiding the teacher to adapt behaviors that the student can feasibly learn.
Three Network Structure.
The approach utilizes three interconnected networks: the teacher, student, and a proxy student, sharing the same action decoder to enhance performance alignment.
Maze Navigation Task Outcome.
In testing, the framework effectively trained a teacher and student to navigate a maze, with the student successfully reaching the goal using learned behaviors.
Agile Quadrotor Control Application.
In quadrotor obstacle avoidance, the approach achieved the highest success rate, with the student effectively adapting its viewing direction to maintain visibility of obstacles.
Robotic Manipulation Improvement.
The method improved success rates by 177% in a drawer opening task, demonstrating the effectiveness of aligned learning in complex robotic scenarios.
Real-World Task Potential.
The framework extends imitation learning capabilities to address multi-modal tasks with significant information gaps, thereby enhancing robotic task performance.
source UZH Robotics
