OG-SPR Improves Visual RL Sample Efficiency.

A new reinforcement learning (RL) algorithm, Observation-Grounded Self-Predictive Representations (OG-SPR), has been introduced to tackle the challenge of sample-efficient policy learning from pixel data in visual continuous control. Existing methods often rely solely on either latent space self-prediction or observation space prediction, which can be insufficient for complex tasks with limited training data. OG-SPR integrates both objectives to learn representations that are simultaneously temporally predictive in latent space and firmly grounded in observation-level dynamics. The algorithm incorporates two key auxiliary objectives: multi-step latent self-prediction and next-observation prediction. To prevent over-constraining the shared representation, OG-SPR uses lightweight adapters for latent self-prediction, allowing the representation to benefit from temporal signals without being forced to directly satisfy the self-prediction objective. Empirical evaluations across 28 visual control tasks from the DeepMind Control Suite demonstrate that OG-SPR outperforms state-of-the-art self-predictive and observation-predictive RL methods, showing notable gains in challenging domains like dog and humanoid locomotion.