Summary

LaST-R1 is an RL post-training framework for reasoning-before-acting VLA policies, introducing Latent-to-Action Policy Optimization (LAPO). LAPO jointly optimizes the latent reasoning process and action generation by embedding latent Chain-of-Thought within the RL loop. An adaptive latent CoT mechanism dynamically modulates the reasoning horizon based on environment state complexity, achieving near-perfect 99.9% average success on the LIBERO benchmark.

Key Contributions

  • Latent-to-Action Policy Optimization (LAPO): jointly optimizes latent CoT reasoning and action generation
  • Adaptive latent CoT: dynamically adjusts reasoning depth based on environment state
  • RL training loop integrates reasoning quality signals alongside task reward
  • 99.9% average success rate on LIBERO benchmark
  • Collaboration between CUHK, PKU, and Simplexity Robotics

Significance

LaST-R1 demonstrates that RL over latent reasoning representations (not just actions) enables near-perfect manipulation performance, suggesting that chain-of-thought reasoning is as valuable in robot control as in language tasks.