VLA lineage
RT-2 · OpenVLA · π₀ · π₀.₅
From discrete action tokens to flow matching, then to long-horizon cleanup in unseen homes. Most later humanoid VLAs branch from this line.
Reading
Notes on other people’s work, mainly humanoid VLA and world models, 2024–2026. These are not my papers.
Start here
Humanoid foundation models
GR00T N1 · Helix · Gemini Robotics
Slow VLM plus a fast action expert. Helix is a company blog, not a paper. GR00T is the open, humanoid-targeted stack.
Whole-body / loco-manipulation
LeVERB · WholeBodyVLA · TWIST
The interface moves from hand-written end-effector commands to learned latents. Large-space loco-manipulation only became its own paper in late 2025.
World models
Cosmos · DreamGen · Cosmos 3 · MotionWAM
World models first became a data factory, then in 2026 some work treats the world model as the policy. MotionWAM is one of the few WAMs aimed at real-time humanoid loco-manipulation.
Also watching
2026
Cosmos Policy · DreamZero · HumanoidArena
Fine-tune a video world model into a policy without changing the architecture. HumanoidArena scores the VLA–tracker interface, not just tabletop success.
Data / fleet
AgiBot World / GO-1
A public humanoid-fleet dataset. ViLLA learns a latent planner from action-free video; WholeBodyVLA still uses that idea.