Post by Hojune Kim
Co-Founder @Dentronic | AeroAstro MS @Stanford | ex-DLR | Robotics
What if a single phone scan was all you needed to teach a humanoid a new skill that still works in scenes it has never seen? Excited to share 🦿LEGS🦿, a photorealistic simulator for training humanoid loco-manipulation with zero teleoperation. Data is the bottleneck for humanoid robots: teleop is slow, keeps the robot in the loop, and does not transfer when the scene or objects change. Here is what LEGS does: 📱 Starts from one handheld phone scan, no teleop and no human demos ⚙️ Places robot and object meshes in MuJoCo for physics, composited over a 3D Gaussian Splatting reconstruction for photorealism 🔁 Re-renders one recorded motion into any new scene, object, or prompt for a fraction of a GPU-hour, with no new demos ✅ Trains policies that deploy zero-shot on a real Unitree G1, matching or beating human teleoperation across three VLA backbones Huge thanks to my collaborators Timothy Chen, Jiankai Sun, Lars Osterberg, Qianzhong Chen, Ke Wang, and Mac Schwager. 📄 Paper: https://lnkd.in/gQHNa-N3 🌐 Website: https://legsvla.github.io
Video Content