LaST-R1: New Physical Reasoning Paradigm Achieves 99.9% Success Rate on LIBERO Benchmark

Researchers from Simplexity Robotics, Peking University, and CUHK introduced LaST-R1, a latent-space physical reasoning framework for robotic manipulation.

The method uses LAPO to jointly optimize reasoning and action, aiming to embed physical reasoning directly in latent space.

LaST-R1 reported 99.9% success on the LIBERO benchmark and a 22.5% improvement over π0.5 in real-world tasks.

The result points to a new path for robot generalization, relying less on language-style chain-of-thought or memorized action patterns.