​D-Robotics Frontier

​D-Robotics Frontier is a technical series where we share the latest advances, research, and engineering breakthroughs from D-Robotics.

​Through this series, we share cutting-edge technologies, research progress, and engineering breakthroughs from D-Robotics, spanning Embodied AI, world models, robot learning, and intelligent robotics.

​More than a showcase of what we’ve built, Frontier is a space for deep technical exchange — bringing together researchers, engineers, developers, and robotics enthusiasts to explore new ideas, challenge assumptions, and shape what comes next in robotics.

​# 2 WM-LOCO

​Today, humanoid robots have become quite good at walking on flat ground, slopes, and various uneven surfaces. But when the terrain becomes more challenging, such as stairs, stepping stones, wide gaps, or narrow balance beams, things get much harder.

​To address this challenge, we present World-Model-Augmented Visual Locomotion (WM-LOCO), which augments a visual locomotion policy with predictive context from a jointly trained world model.

​On November 19, we’re hosting a technical webinar to open up the technology behind WM-LOCO.

• ​The Challenge: Foothold-Constrained Locomotion

• ​Our Approach: Joint World Model and Locomotion Policy Learning

• ​Performance Results

  • ​Simulation success rates: 96.9% / 96.0% / 94.0% / 92.1% across the four terrain types

  • ​PPO baseline: 0% success rate on stepping stones, wide gaps, and balance beams without a world model

  • ​Stair locomotion: 25.6% increase in stride length and 28.9% reduction in pelvis acceleration

• ​Real-World Deployment & Validation

  • ​Direct sim-to-real deployment on Unitree G1, with no additional real-world training

  • ​37 successful trials out of 40 (92.5%) across four terrain types

  • ​Perception based on visual input and proprioception

  • ​1.5-second action-conditioned prediction: predicted depth frames track the actual scene, while different action sequences produce increasing prediction errors

​📅 Nov 19, 21:00 GMT+8

​🎙️ Speaker: Yuxi Liu | Algorithm Engineer at D-Robotics

​Speaker bio: Yuxi Liu is a robotics researcher at D-Robotics, where he is responsible for humanoid locomotion research. He received his M.S. in Robotics from Columbia University. His research focuses on reinforcement learning, world models, and learning-based control for humanoid robots, with an emphasis on robust locomotion in challenging environments. He is dedicated to advancing robust and adaptive whole-body control that integrates locomotion and manipulation, enabling humanoid robots to move, interact, and perform useful tasks reliably in the real world.