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

A joint research team from Simplexity Robotics, Peking University, and the Chinese University of Hong Kong has proposed the LaST-R1 model, a new paradigm for integrated AI training. Published as a prominent paper at ICML 2026, LaST-R1 achieved a 99.9% success rate on the LIBERO benchmark with just one warm-up batch, effectively setting a new standard. LaST-R1 surpassed the previous SOTA model π0.5 by 22.5% in real-world tasks such as grasping and rotation.