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.

• →MIKASA provides the first standardized benchmark for assessing memory capabilities in reinforcement learning across diverse scenarios.
• →The suite includes MIKASA-Robo with 32 carefully designed tasks specifically for tabletop robotic manipulation.
• →Memory is identified as crucial for solving complex tasks with temporal and spatial dependencies in robotics.
• →The benchmark addresses a significant gap in the field where no universal memory assessment tools previously existed.
• →The framework aims to advance memory RL research and enable more robust real-world robotic systems.