Jin Cheng
240 posts
@catachiii
Doctoral Student at @crl_ethz @ETH, working on RL, robotics, and more - Cells interlinked within cells, interlinked.
We recently explored how to learn a time-varying linear (TVL) policy for character control. It works surprisingly well! In simulation, a TVL policy can handle every Deepmimic-style task we throw at it. No neural net at deployment, just a sequence of matrices.
Introducing our recent work: 🐶Teaching Robots Like Dogs: Learning Agile Navigation from Luring, Gesture, and Speech We ask a simple question: Can we teach robots the way we teach dogs? Project page: lnkd.in/eWWSJds2
NEO The Home Robot Order Today
I've long wondered if we can make a humanoid robot do a 𝘄𝗮𝗹𝗹𝗳𝗹𝗶𝗽 - and we just made it happen by leveraging 𝗢𝗺𝗻𝗶𝗥𝗲𝘁𝗮𝗿𝗴𝗲𝘁 with BeyondMimic tracking! This came after our original OmniRetarget experiments, with only minor tweaks to RL training: relaxing a termination threshold and removing one reward term. The policy achieved a 𝟱/𝟱 success rate in our real-world experiments, showing the strength of high-quality, interaction-preserving motion retargeting combined with BeyondMimic’s minimal RL tracking. Here is the updated arXiv: arxiv.org/abs/2509.26633 (In Sec. V. A)
