Gasgoo Munich- Gasgoo Embodied Intelligence has learned that D-Robotics recently unveiled WM-LOCO, a locomotion solution designed for humanoid robots navigating complex terrain. The system marks the first end-to-end joint training of a predictive world model with a PPO reinforcement learning policy, aimed at sharpening robots' ability to walk continuously and make decisions where footholds are limited.

Image Source: D-Robotics
On the technical side, WM-LOCO skips foothold annotation, teacher distillation, and multi-module serial pipelines — training directly from visual input to joint control commands in a single step. Its recurrent world model fuses onboard depth images, proprioceptive signals, and action history to learn how movements shape subsequent body states and visual observations. The predictive internal features generated by the world model feed directly into control, allowing the locomotion policy to coordinate body motion with foot placement without first rolling out multi-step predictions and searching for an optimal path.
Under identical sensory inputs and training budgets, adding the world model markedly improved the robot's ability to traverse plum blossom piles, trenches, and balance beams. In simulation, WM-LOCO achieved a near-100% pass rate across four types of discrete sparse terrain — plum blossom piles, stairs, trenches, and balance beams. When the same trained weights were deployed zero-shot onto a Unitree G1 physical robot, testing across all four terrain types (10 trials each) yielded an overall success rate of 92.5%. The robot also cleared challenges including a 0.8-meter-wide trench and a 10-centimeter-wide balance beam.
The solution has since been demonstrated in real time at public events on the Sunrise S600 computing platform, validating the algorithm's stability and engineering reliability under real-world sensor noise and physical conditions. Related technical papers and a project homepage have been released publicly.









