Issues running cosmos-reason1 on Jetson AGX orin

Robotics & Edge Computing Jetson & Embedded Systems Jetson AGX Orin
,
1

I started with jetson-containers run $(autotag cosmos-reason1) to run cosmos-reason1. But when I checked container folders, it didn’t download cosmos-reason1 from github GitHub - nvidia-cosmos/cosmos-reason1: Cosmos-Reason1 models understand the physical common sense and generate appropriate embodied decisions in natural language through long chain-of-thought reasoning processes. . After that, inside the jetson-container, I went thorough the cosmos-reaosn1 installation from GitHub - nvidia-cosmos/cosmos-reason1: Cosmos-Reason1 models understand the physical common sense and generate appropriate embodied decisions in natural language through long chain-of-thought reasoning processes. . I faced lots of dependency compatibility issues getting the torch, torchcodec and torchvision installed on the Jetson AGX platform which shows as “manylinux_2_35_aarch64”.

I overcame those issues one by one but when the following script started working,

uv run scripts/inference_sample.py

I got he following error.

Screenshot from 2025-10-27 10-57-56
Screenshot from 2025-10-27 10-57-561552×656 133 KB

My Jetson AGX orin is currently on CUDA 12.9. I am currently looking to just run the cosmos-reason1 model on Jetson AGX in someway but seems like that’s happening because of these dependencies issues.

Has any one being able to run cosmos-reason1 successfully on Jetson AGX orin?

@neel_patel

5

Hi,

Which container do you use?
Is the one below?

https://hub.docker.com/r/dustynv/cosmos/tags

Thanks

6

I used the autotag through the following command jetson-containers run $(autotag cosmos-reason1) and it automatically pulled the following container.

cosmos-reason1:r36.4.tegra-aarch64-cu129-22.04-onnx

Screenshot from 2025-10-27 23-19-57
Screenshot from 2025-10-27 23-19-571850×345 81.4 KB

https://hub.docker.com/r/dustynv/cosmos/tags , This link doesn’t specify container for cosmos-reason1 and cu129. All the containers are for cosmos and not cosmos-reason1 and two of the containers at this link are for cu128.

7

Hi,

The container is not completed.
In jetson-container, the container name should look like cosmos-reason1:r36.4.tegra-aarch64-cu129-22.04.
An extra - means it is the dependencies image, and the target image is not built successfully.

Do you mean any errors when building the container?
Could you try jetson-containers build $(autotag cosmos-reason1) to see if everything is okay?

Thanks.

8

Thanks for the suggestion. I tried the above but got the following error

KeyError: “couldn’t find package: cosmos-reason1:r36.4.tegra-aarch64-cu129-22.04-onnx”

Screenshot:

Screenshot from 2025-10-28 00-15-53
Screenshot from 2025-10-28 00-15-531851×761 115 KB

9

Hi,

Could you share the document you follow so we can know more about which container you needed?

Thanks.

10

I was trying to the run the following github repo inside the Jetson-container for cosmos-reason1

nvidia-cosmos/cosmos-reason1: Cosmos-Reason1 models understand the physical common sense and generate appropriate embodied decisions in natural language through long chain-of-thought reasoning processes.

jetson-containers/packages/diffusion/cosmos/cosmos-reason1 at master · dusty-nv/jetson-containers

11

Hi,

The corresponding tutorial for the Jetson-Container can be found below:

Could you follow the steps to see if it can work?
We didn’t find the inference_sample.py mentioned on the page.

Thanks.

12

The file is in the scripts folder.

cosmos-reason1/scripts at main · nvidia-cosmos/cosmos-reason1

I was looking to run “cosmos-reason1” for reasoning on a video and not “cosmos” for an image/video generation. Can I run ”cosmos-reason1” within “Cosmos” container?

13

Hi,

We only have the cosmos example.
You might need to handle the compatibility issue for cosmos-reason1 for using it.

Here is a related post for your reference:

Thanks.