Unitree Opens UnifoLM-WLA-1.0, One G1 Policy for 64 Tasks
Unitree put a 6-billion-parameter whole-body foundation model on the table today and said it is fully open-sourcing it. UnifoLM-WLA-1.0 is trained on about 2,500 hours of real-robot data. One checkpoint, according to the project page, covers 64 G1 tasks that mix tabletop work with walking, bending, and carrying.
The demo still that matters is a G1 walking toward an open front-loader with green laundry in its left hand, stereo head and wrist feeds on the side, and policy logs sitting under 105 ms inference. That is household motion, not a kung-fu clip.
One checkpoint, two workspaces
Humanoids Daily splits the 64-task eval the way Unitree presents it:
- 10 whole-body tasks: trash, clothes into a washer, bed-making, inventory on shelves
- 54 tabletop tasks: towel folding, conveyor sorting, uncapping, plugging in cords
The same policy is supposed to drive two-finger grippers and more than one five-finger hand. Cross-end-effector transfer is the claim that will get stress-tested first. If swapping a clamp for a dexterous hand does not force a full data recapture, that is the useful part for anyone sitting on a pile of G1s.
The stack has three layers. UnifoLM-ER-1-4B, built on Qwen3-VL-4B, is the embodied reasoner, trained on more than 5 million samples (points, boxes, multi-image QA, 2-D trajectories, 3-D detection) mixed with ordinary image-text data. On Unitree’s own 16-benchmark sheet, ER-1 posts 62.4 spatial understanding versus 47.7 for Qwen3-VL-4B. UnifoLM-ER-Flow then predicts future dynamic-region masks from optical flow, quantized through a VQ-VAE. An MMDiT action expert decodes continuous motion. Residual vector quantization splits the action space into end-effector, hand/gripper, and lower-body streams.
Training data includes Unitree’s open datasets and BitRobot-HIW-500. IT Home, citing the company, repeats the 64-task, cross-task, cross-end-effector pitch and the “SOTA among open models” line. Treat those leaderboard wins as Unitree’s own eval until independent numbers show up.
Open, with a caveat
The project page still labels Code, Models, and Datasets as coming soon. The announcement is real. The download buttons are not live yet. That is the difference between a press drop and a repo you can pip install tonight.
A Human’s Take
I care about the washer clip more than the 6B count. A single policy that walks, stoops, and stuffs clothes is the shape of work people actually want from a G1. I also want the weights, not a landing page that says “coming soon.” Unitree already flooded the market with cheap bodies. If WLA-1.0 ships as promised, the interesting question is whether 2,500 hours is enough to survive a kitchen that is not in the training set.