Motion Mamba 2.0

Motion Mamba 2.0: Autoregressive Diffusion Mamba for Long-Horizon Humanoid Control

1Peking University 2Korea University 3Monash University 4Nanyang Technological University

∗Equal contribution. †Project lead. ‡Corresponding author.

Overview

One model. Endless motion.
From text prompts to humanoid robots in the real world.

Autoregressive motion generation. Long-horizon control.
Diverse behaviors in one continuous stream.

Simulation

Generated motion. Physics-driven execution.
One synchronized stream of movement.

Real World Demonstration

Long-horizon motion. Changing instructions.
One continuous stream of movement.

Real Viser
Real Viser
Real Viser
Real Viser

Method

Overview of the Motion Mamba 2.0 framework
Overview of Motion Mamba 2.0. Conditioned on text, spatial controls, and accepted motion history, the framework generates motion autoregressively through a two-stage diffusion denoiser. At each denoising step, the root network predicts the global trajectory, whose local features condition body prediction. Both networks combine attention and Mamba-3 through a hybrid sequence mixer. After each clean motion window is accepted, the history states are updated for subsequent generation.
Streaming generation and hybrid attention–Mamba mixer
Streaming generation and hybrid attention–Mamba mixing. (a) Motion is generated window by window through root–body diffusion, with the history held fixed during denoising and updated only after the sampled window is accepted. (b) Base attention retrieves recent motion features, while PCM attention additionally accesses persistent content slots. Their gated residual difference is combined with the Mamba-3 output resumed from the committed recurrent state.
CUDA Graph replay and fused Mamba transition kernel
CUDA Graph and Fusion Kernel. CUDA Graph replay of the three CFG branches and fused Mamba transition preprocessing.