WALL-WM is our latest effort on building foundation world-action-models for general purpose robotic manipulation, using “semantically coherent action events” as the unit of world-model learning.

IMPACT learns an internal model of the environment dynamics within the controller and compensates for these dynamics in a feedforward manner.

OAT enables scalable autoregressive robot control by converting continuous actions into compact, causally ordered tokens.

SIMPACT equips Vision-Language Models (VLMs) with physical reasoning through simulation-in-the-loop world modeling.

The Seed2.0 series models are our latest effort to build foundation models that meet the requirements of large-scale production deployment.

ASAP enables agile, whole-body humanoid motion control by using data to align simulation with real-world physics.

CooHOI is a framework for cooperative object transportation tasks involving multiple humanoid characters, inspired by how humans learn to collaborate through implicit communication.

HIMLoco enables quadruped robots to adapt to various terrains by learning to estimate environmental disturbances online both explicitly and implicitly.

FamO2O is a framework for offline-to-online reinforcement learning algorithms that determines state-adaptive improvement-constraint balances.