cyclotron

Building an Open-Source Humanoid | Yip Jia Qi, Menlo Research | Φ Podcast Ep 2 What does it take to build a humanoid that anyone can download, repair and change? Φ sat down with Yip Jia Qi, lead research engineer at Menlo Research, the lab behind Asimov, an open-source humanoid, and the Kepler encoder. Jia Qi came to robots from PhD research in speech and audio. He joined Menlo to put a voice agent inside a robot, and first had to help build the robot. We get into why Asimov is open source, what zero-shot sim-to-real really took, how Cyclotron trains walking with AMP, robots that improve their own skills and their own bodies, imitation learning versus RL, Kepler and the case for one shared encoder, why touch is so hard, and how you test and deploy a humanoid. Then rapid fire, and three predictions we will check next year. Hosted by Yashvardhan Gupta and Parv Patodia. CHAPTERS 0:00 Intro 0:36 From speech research to building robots 3:07 Asimov, and why open-source a humanoid 6:21 Zero-shot sim-to-real: it comes down to the actuator 16:25 Cyclotron, AMP vs PPO, and why RL feels like gambling 20:53 Vision: stairs, LiDAR, and a head you can replace 29:11 Self-improving robots: new skills, then new bodies 36:19 Imitation learning vs RL: the final boss 40:54 Kepler: one encoder, many senses, many models 50:04 Why force and touch are hard to give a robot 54:09 Asymmetric cross-attention, the Perceiver, and JEPA 1:00:44 Evaluation: a fitness test for robots 1:03:40 Deployment: why onboarding is the data that matters 1:07:39 Can a robot hear that something is wrong? 1:10:33 Rapid fire 1:16:15 Predictions: robots in homes, and robots that pay tax 1:19:06 Can we solve robotics this time? MENTIONED IN THIS EPISODE Asimov 1, the open-source humanoid: menloresearch/asimov-1 How we achieved zero-shot sim2real for Asimov: menlo.ai Noise is all you need to bridge the sim-to-real locomotion gap: menlo.ai Cyclotron: menloresearch/cyclotron Menlo SDK: menloresearch/menlo-sdk AMP, Adversarial Motion Priors: arxiv.org Kimodo: research.nvidia.com SONIC: nvlabs.github.io Kepler-Encoder v0.1: arxiv.org ACA-Net: arxiv.org Perceiver: arxiv.org Perceiver IO: arxiv.org LeJEPA (SIGReg): arxiv.org MolmoBot: arxiv.org GUEST Yip Jia Qi: linkedin.com ABOUT Φ Φ (Physical Hardware Intelligence) is a robot learning lab in Silicon Valley. We build on the SO-101 arm and share everything we learn. physical-hardware-intelligence.github.io

@notcalmerecentl hey, thanks for sharing it with us! We're on it. Feel free to start with this: menloresearch/cyclotron A platform with a leaderboard is next.

You can train Asimov 1 using our open-source RL training framework, built on @NVIDIAAI's Isaac Lab. menloresearch/cyclotron

@rosario_scaling @octi_zhang asimov! menloresearch/cyclotron thats the one I'm training right now from @asimovinc

@NRaghav_4 @asimovinc mjlab: menloresearch/cyclotron and isaac sim (coming soon) you can also try it here: try.menlo.ai

You can start training a humanoid before you own one. Asimov 1's open-source training code gives you a foundation to train and evaluate locomotion policies in simulation while you build your robot or wait for it to arrive. Train your own policy: menloresearch/cyclotron

You can start training a humanoid before you own one. Asimov 1's open-source training code gives you a foundation to train and evaluate locomotion policies in simulation. Train your own policy: menloresearch/cyclotron t.co