Reflex: Exploiting Reflection Symmetry in State-Based RL
A new reinforcement learning paradigm called Reflex exploits reflection symmetry in state-based continuous control tasks to improve sample efficiency. The work formalizes axial and bilateral reflection types and their transformations, integrating symmetry into policy learning via regularization mechanisms. Reflex is compatible with both on-policy and off-policy algorithms. The approach addresses the underexplored area of reflection symmetry in state-based RL, contrasting with prior focus on image-based RL and rotational symmetry like SO(2). The paper is published on arXiv under identifier 2605.23415.
Key facts
- Reflex is a paradigm for reinforcement learning with reflection symmetry.
- It targets state-based continuous control tasks.
- Two types of reflection are formalized: axial and bilateral.
- Reflex integrates symmetry via principled regularization mechanisms.
- It works with both on-policy and off-policy RL algorithms.
- Prior work focused on image-based RL and rotational symmetry (SO(2)).
- The paper is on arXiv with ID 2605.23415.
- Reflex aims to improve sample efficiency in RL.
Entities
Institutions
- arXiv