Mondo Robotics

“A team from Mondo Robotics and HKUST has cracked a core bottleneck in humanoid robotics: getting world-model-based policies to run fast enough for real-time whole-body control.”

“The result is a 30%+ performance jump over the best VLA baselines on tasks that require coordinated full-body movement — including leg-driven behaviors that hierarchical systems literally cannot execute.”

“A Video DiT initialized from Cosmos-Predict2.5-2B — a causal spatio-temporal VAE plus a flow-matching diffusion transformer conditioned on Cosmos-Reason1 language embeddings.”

“Even π_0.5 stays under 20% overall, indicating that strong semantic priors alone do not transfer to the closed-loop physics humanoid loco-manipulation demands.”

“We conduct all real-world experiments on a Unitree G1 humanoid with dual ALOHA2 grippers.”

“GR00T-X-Embodiment-Sim (GR1): Fourier GR1 (sim), weight 0.255 and GR00T-Teleop-GR1-Robot: Fourier GR1 (real), weight 0.071 — combined 32.6% of total Stage 1 budget from a single non-target embodiment.”

“We instantiate m_t on top of SONIC [4], a universal whole-body controller that fuses different motion targets through a single shared latent.”

“Ma et al. [2026] T. Ma, J. Zheng, Z. Wang, C. Jiang, A. Cui, J. Liang, and S. Yang. DiT4DiT: Jointly modeling video dynamics and actions for generalizable robot control.”

“∗ Corresponding author, Co-advising.”