Chris Paxton posts humanoid agility work

- Robotics researcher Chris Paxton shared results using flow‑matching plus RL residuals to train humanoid agility policies, noting dozens of robots were broken during development. (x.com) - The methods combine diffusion/flow matching generative trajectory priors with small RL corrections to close gaps between simulation and reality, applied to humanoid platforms. (x.com) - The post signals that agility at scale still needs hybrid pipelines and more robust hardware-testing regimes as labs push policies into real bodies. (x.com)