Unitree G1 - Whole-body VLA
From VR teleoperation to an autonomous, language-conditioned VLA policy on the Unitree G1 humanoid: teleop, IK + RL vs. SONIC, data collection, and a deployed Psi0 policy.
Technical Notes
Blog
Technical notes, and mathematical reflections on AI and Robotics.
The Math Behind World Model Paradigms for Robot Learning
A mathematical analysis of IDM-style, Single-backbone-style, and MoT-style world-model paradigms in robot learning from the perspectives of probabilistic modeling, structured optimization, gradient conflict, and parameter coupling.