Reproduce “TraKDis: A Transformer-based Knowledge Distillation Approach for Visual Reinforcement Learning with Application to Cloth Manipulation” 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

Wei Chen, Nicolas Rojas

Imperial College London researchers developed TraKDis, a Transformer-based knowledge distillation approach, enabling robots to perform complex cloth manipulation using only visual input. This framework achieved up to 21.9% higher performance compared to existing visual reinforcement learning methods in simulation and demonstrated effective real-world transfer.

Key insights

  • Re-framing RL as sequence modeling with Decision Transformers enables agents to leverage long-term historical information, which improves control in dynamic, partially observable environments like cloth manipulation.
  • Knowledge distillation is highly effective when the student policy's weights are initialized from a pre-trained privileged teacher, drastically reducing training time and improving the final performance of the visual agent.
  • Robust state estimation from visual data, achieved through a pre-trained CNN encoder and image augmentation, is critical for successfully distilling knowledge from a state-based teacher to a vision-based student.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.