Memory, Benchmark & Robots: A Benchmark for Solving Complex Tasks with Reinforcement Learning

Researchers introduce MIKASA, a comprehensive benchmark suite designed to evaluate memory capabilities in reinforcement learning agents, particularly for robotic manipulation tasks. The framework includes MIKASA-Base for general memory RL evaluation and MIKASA-Robo with 32 specialized tasks for tabletop robotic manipulation scenarios.