D-Robotics Unveils WM-LOCO Locomotion for Humanoids
The end-to-end system takes visual input directly to joint control for humanoid robots navigating complex terrain.
Updated on Nov. 20, 2026 in Robotics

D-Robotics has unveiled WM-LOCO, a locomotion solution for humanoid robots that was demonstrated in real time on its Sunrise S600 computing platform. The system is described as the first end-to-end joint training of a predictive world model with a PPO reinforcement learning policy.
Why it matters
WM-LOCO skips foothold annotation, teacher distillation, and multi-module serial pipelines, trading that structure for a single step from camera input to joint commands. Related technical papers and a project homepage have been released publicly, allowing outside researchers to examine the approach.
In simulation, WM-LOCO reached a near-100% pass rate across plum blossom piles, stairs, trenches, and balance beams, and on a physical Unitree G1 it achieved 92.5% success over 40 trials. The robot cleared a 0.8-meter trench and a 10-centimeter balance beam, though the source does not address performance on terrain beyond these four types.
The players
D-Robotics
D-Robotics is the developer of the WM-LOCO locomotion system and the Sunrise S600 computing platform on which it was demonstrated.
Unitree G1
The Unitree G1 is a physical humanoid robot on which WM-LOCO was deployed zero-shot to achieve its reported 92.5% success rate.
The details
WM-LOCO fuses onboard depth images, proprioceptive signals, and action history in a recurrent world model that feeds predictive features directly into the locomotion control policy. The policy coordinates body motion with foot placement without first rolling out multi-step predictions.
Timeline
Article on the WM-LOCO unveiling was published on 2026-11-20.
The Tech Race
WM-LOCO extends PPO reinforcement learning, a widely used method for training robot control policies, by jointly training it with a predictive world model in a single pipeline. This departs from multi-module serial approaches that separate perception, teacher distillation, and foothold annotation into distinct stages.
For developers and researchers building humanoid robots, WM-LOCO offers a locomotion approach that works directly from depth images and joint feedback without hand-annotated footholds. Its zero-shot transfer to physical hardware means teams may test the approach on existing robots without retraining from scratch, though availability beyond published papers is not described in the source.
The takeaway
Locomotion control for humanoid robots is moving toward training perception and movement together rather than in separate stages. Readers following robotics can track whether the reported results hold on additional terrain and hardware.
Further reading
For related coverage of machines that move through the physical world, see Robotics on Wisevoter.
Source note: This article includes information reported by Gasgoo.



