A motion-token paper gives a Unitree G1 longer moves to follow. It is still research.

The result

A new paper from researchers at Technical University of Darmstadt reports a way to turn longer humanoid movements into compact tokens that a robot controller can follow. The authors tested it on a physical Unitree G1. Their reported demonstrations include generated walking, dancing and fighting motions, plus a 35-second transition from fighting to dancing.

That is a real-robot control result, not a consumer product announcement. The paper was submitted to arXiv on October 27 and says it was accepted to Humanoids 2026. Its claim is narrower than a household robot: a tracking policy follows a detailed future motion reference better when that reference is represented in learned chunks rather than as a simple velocity command.

The authors compare their stochastic tokenizer with finite scalar quantization and VQ-VAE at matched capacity. They report that the chosen representation reconstructs G1 motion better in their evaluation. The public arXiv listing for the new robotics papers independently lists the submission and its Humanoids 2026 note.

How it works

The system has three layers. First, it learns a compact representation of whole-body motion from recorded trajectories. Second, an autoregressive model generates a sequence of those representations from a behavior label or text command. Third, a reinforcement-learning tracking controller uses the sequence as a detailed reference for the robot.

The division matters. A conventional velocity command says roughly where to go; it does not supply much information about how a whole body should prepare for a turn, a gesture or a change between activities. The authors argue that a longer motion representation gives the controller enough look-ahead to make those transitions.

The experiment is still bounded. The paper describes physical G1 motion execution, not a general language-controlled assistant. Its examples are locomotion and expressive motion. There is no reported test of a robot recognizing a household object, choosing a task plan, recovering from clutter, or safely working around people.

What it means for a home robot

A home humanoid needs its arms, torso and legs to cooperate while it opens a cabinet, carries a load or changes direction around furniture. Work that produces more stable and controllable whole-body movement is therefore useful infrastructure. A controller that can change behavior without restarting from a fixed clip could eventually make a robot less brittle in those transitions.

But motion quality is one layer of the stack. A useful home robot also needs perception, grasping, force control, task planning, reliable recovery and safety constraints. This paper does not claim to solve those layers. The G1 demonstrations establish that the learned representation can be used on hardware for the stated motions, not that the robot can complete domestic work.

For owners, the near-term significance is a clearer direction of travel: researchers are trying to give humanoids richer motion references than a single command at a time. That is not the same as autonomy. It may help a future machine move more gracefully after it already knows what to do.

Caveats

The evidence comes from the authors' paper and the reported experimental setup. The work has not been presented here as an independent product test, and the paper does not report customer use, a home deployment or a benchmark of household chores. Its physical demonstrations involve one hardware platform, the Unitree G1.

The source says the paper is accepted to Humanoids 2026, but an arXiv posting remains a research disclosure rather than a shipping commitment. Treat the result as a specific control technique demonstrated on a real humanoid, not as evidence that a G1 or any other humanoid is ready to work independently in a home.