Reproduce “Student-Informed Teacher Training” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Nico Messikommer, Jiaxu Xing, Elie Aljalbout, Davide Scaramuzza
The Robotics and Perception Group at the University of Zurich developed a "Student-Informed Teacher Training" framework that modifies a reinforcement learning teacher's objective to learn behaviors that are inherently more imitatable by a partially observable student. This approach significantly increases student success rates in tasks like quadrotor obstacle avoidance and robotic manipulation by encouraging perception-aware teacher policies.
Key insights
- Actively shaping the teacher's policy to be "student-aware" during its training is more effective than solely improving the student's imitation capabilities.
- Modifying the teacher's objective can lead to the emergence of desirable, student-friendly behaviors, such as perception-aware navigation or self-occlusion avoidance, without explicit reward engineering.
- Aligning the teacher's policy with the student's capabilities at the source not only improves student performance but can also lead to more robust and efficient teacher policies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.