Open research on humanoid and legged robot learning, from Vietnam.

Robotics Research Hub Vietnam builds and shares open-source tools for training controllers for humanoid and legged robots. Our work centers on fast, GPU-native robot learning: differentiable world models, adversarial motion priors, and reinforcement learning that runs end-to-end on a single GPU with JAX and MuJoCo Warp. We aim to make state-of-the-art robot learning reproducible and affordable for students and researchers in Vietnam and beyond.

• Humanoid locomotion: natural walking, running and get-up behaviors for platforms such as the Unitree G1
• Robot learning: reinforcement learning, imitation from human motion, and adversarial motion priors (AMP)
• Differentiable simulation and world models: policy optimization with analytic gradients through learned dynamics
• Efficient training: fully compiled pipelines that train in minutes on one consumer GPU

We welcome issues, pull requests and research collaborations. If you use our code in your work, please cite the corresponding repository.