Humanoid foundation modelsUnitree

UnifoLM-WLA-1.0: what its action interface means for robot integration

Read Unitree's released action/state layout: masked hardware slots, action-chunk timing and different leg offsets in the state and action vectors.

By TechniaHQRobot2 min read
Unitree G1 humanoid robot

Introduction

Unitree has published the repository for UnifoLM-WLA-1.0. The company describes it as a 6B-parameter general-purpose humanoid foundation model trained with about 2,500 hours of real-robot data.

The reported task set covers 64 tabletop and whole-body manipulation tasks. The same model is designed to work with two-finger grippers and several five-finger hands.

Source review date recorded for this article September 29, 2026.

The action vector is an interface, not a joint count

The released processing guide defines 54 action values and 60 state values. Slots cover different end effectors and body modules; unavailable slots are zero-filled and masked. A 54-dimensional output therefore does not mean every supported robot has 54 actuated joints. Cross-hand support describes a common representation, not equal performance across hands.

The open release is split into pieces

The repository records the training and fine-tuning code and model release on September 28, 2026. LoRA support remains planned. Its processing guide resamples actions at 30 Hz and uses 30-point chunks, spanning roughly one second; this is an action horizon, not a measured motor-control frequency.

A state/action indexing trap

The guide places left and right leg actions at indices 42–47 and 48–53, but leg states at 48–53 and 54–59. Reusing the state offsets for the action vector would map the wrong channels. Check the two layouts separately when adapting a dataset.

The projector expects a state mask, not sixty additional measured joints

Unitree's public processing specification uses a 60-dimensional physical state vector and a separate 60-dimensional validity mask. When the robot-state projector is active, those two arrays are concatenated to create a 120-dimensional input. The second half indicates which fields the embodiment provides; it does not describe sixty extra actuators. Reusing the action mask in this position would also be incorrect, because the action representation has its own 54-dimensional layout.

A developer adapting a dataset should validate the source frame, pose rotation encoding, leg-slice offsets, normalization statistics and action-chunk frequency before inference. The 30-frame chunk at a 30 Hz target describes the data sequence, not an independently measured 30 Hz motor command loop. A reproducible evaluation needs the adapter code, exact hand and torso configuration, and a complete task log with missed actions and resets.

Sources and methodology

UnifoLM-WLA-1.0 repositoryUnitree Robotics · Accessed September 29, 2026Robot action/state processingUnitree Robotics · Accessed September 29, 2026
Article by @techniahqrobot