From：Internet Info Agency 2026-09-08 09:08:00
Recently, Digant Robotics unveiled its WM-LOCO framework, which for the first time integrates a predictive world model into humanoid robots for complex terrain locomotion and enables end-to-end joint training with a PPO policy. This approach eliminates reliance on footstep annotations, teacher distillation, or multi-module sequential pipelines, directly generating joint control commands from visual inputs in a single training step. WM-LOCO undergoes end-to-end joint training entirely in simulation and achieves zero-shot real-world transfer—meaning policies trained in simulation can be deployed directly onto physical robots without any fine-tuning. The system simultaneously processes onboard depth vision and proprioceptive data, leveraging a recurrent world model to learn dynamic relationships among the robot, its actions, and the environment within a latent space. In real-world tests on the Unitree G1 robot, the same set of weights was deployed in a zero-shot manner across three terrain types—stepping stones, stairs, and gaps—with 10 trials each, achieving an average success rate of 93.3%. Notably, the robot successfully cleared a 0.8-meter-wide gap in a single stride, executing a complete “swing-leg–cross–landing” motion. Wider gaps were not tested due to safety considerations. The framework has already been publicly demonstrated on the Sunrise S600 computing platform, where the algorithm runs in real time on a physical robot equipped with real sensors and the S600 platform, closing the full loop from visual perception and world modeling to whole-body motion control.
