Goose tidies a house using off-the-shelf open models

Humphrey Hu posted a robot called Goose tidying a home after a user asked it to make the place safe for a visiting one-year-old. The framing is the interesting part. Picking things up is one problem, deciding which objects are dangerous for a toddler is a different one, and Hu says the robot is doing both.

The stack is the headline. Hu says Goose runs Qwen3-8B as the slow reasoning layer on an RTX 5090, and NVIDIA's GR00T n1.7 as the fast control layer, a VLA (a model that turns camera images and text into robot actions), running onboard a Jetson AGX Thor. Both are open models anyone can download. Nothing in the post is a bespoke in-house foundation model.

The claim that got attention

Hu's pitch is about cost and time, not capability. He says this kind of behavior used to need a team of researchers, a billion dollars and a few years, and now takes weeks with open source models if you know what you are doing. His closing line is that knowing how to build something people actually love is the part that does not go out of style.

Ryan Julian, replying to the clip, said his timeline for home robots has moved up. Chris Paxton went the other way and questioned the architecture itself, saying he is still waiting for something like a Qwen 3.8 at 8B or 16B and that he is not sure you need the VLA for pick and place tasks at all. That is a real disagreement about where the hard part lives. If a general language model with a good grasp of the scene can drive simple grasps directly, the fast policy underneath becomes a smaller piece of the puzzle than the current two-system framing suggests.

Shown versus claimed

What is shown is a tidying run on video with a named stack behind it. What is not stated anywhere in Hu's post is how much of the run was autonomous, how long it took in real time, whether the footage was sped up, or how many attempts it took to get a clean one. Hu lists the models and the hardware and leaves the rest alone.

So treat the hardware claim as solid and the weeks versus years claim as Hu's own description of his own project. Nobody outside has reproduced it yet. The part that is checkable and genuinely new is that two publicly available models, one for reasoning and one for control, were enough to put a household task on video at all.

Ryan Julian

@ryancjulian

X

my timeline for home robots has definitely moved up

Chris Paxton

@chris_j_paxton

X

Im not sure if you will need the vla for pick and place tasks.

Humphrey Hu

@thehumhuworks

X

Now it takes week(s) with open source models (if you know what you're doing).