Skip to content
embodied.guide

Back to the AI models

NVIDIAUnited States2026

Cosmos Policy

An AI model: a vision-language-action model (VLA) takes camera images and an instruction and turns them into movements for a robot. This page describes the software, not a robot.

Robotsnot stated

Key figures as stated by the developer

5 of 6 stated · 1 gap
  • Model type not stated
  • Controls ⁠A arms and hands
  • Inputs ⁠A camera language robot state
  • Access ⁠B open download
  • Parameters ⁠A 2,000,000,000
  • Base model ⁠A NVIDIA Cosmos-Predict2-2B-Video2World
How to read the figures
14kgA30.9lbA
The letter in the grey square after a figure points to the source the figure came from. Robots with several sources carry several letters.
74.5lb
Imperial figures are our own conversion unless the manufacturer states an imperial figure itself. The source letter always belongs to the metric figure — a conversion has no source of its own.

The four states

Versions and notes in the key figures (6)

Model type

Note The developer does not label it vision-language-action, world action model or embodied reasoning: "is a 2B-parameter bimanual robot manipulation policy model fine-tuned from the NVIDIA Cosmos-Predict2-2B-Video2World video foundation model.".

Controls

Model card (ALOHA checkpoint): "Action chunk: 50-timestep sequence of 14-dimensional actions (7 per arm: joint positions for 6 joints + 1 gripper)"

Inputs

Model card (ALOHA checkpoint): "Input Type(s) : Text + Multi-view Images + Proprioceptive State"

Access

Paper abstract: "We release code, models, and training data". "Hugging Face [01/22/2026] via https://huggingface.co/collections/nvidia/cosmos-policy"

Parameters

Model card: "2B (inherited from base model)"

Base model

"Base model : NVIDIA Cosmos-Predict2-2B-Video2World"

Full specifications

9 of 12 fields stated
The letter points to the source at the foot of the page. Fields without a figure are dashed: the developer does not state them — that is not the same as zero, and not the same as no.
Field Value Source and caveats
Input and output
Model typenot stated

Note The developer does not label it vision-language-action, world action model or embodied reasoning: "is a 2B-parameter bimanual robot manipulation policy model fine-tuned from the NVIDIA Cosmos-Predict2-2B-Video2World video foundation model.". ⁠A

Inputscamera language robot stateA

Model card (ALOHA checkpoint): "Input Type(s) : Text + Multi-view Images + Proprioceptive State"

Controlsarms and handsA

Model card (ALOHA checkpoint): "Action chunk: 50-timestep sequence of 14-dimensional actions (7 per arm: joint positions for 6 joints + 1 gripper)"

Degrees of freedom controllednot stated

Note Action size for the ALOHA 2 robot with two ViperX 300 S arms: "Action chunk: 50-timestep sequence of 14-dimensional actions (7 per arm: joint positions for 6 joints + 1 gripper)" ⁠A
Size and speed
Parameters2,000,000,000A

Model card: "2B (inherited from base model)"

Control rate25HzA

Rate of the robot's actions (ALOHA checkpoint): "Control frequency: 25 Hz (reduced from original 50 Hz)"
Where it runs and what it is built on
Runs wherenot stated
Base modelNVIDIA Cosmos-Predict2-2B-Video2WorldA

"Base model : NVIDIA Cosmos-Predict2-2B-Video2World"
Openness
Accessopen downloadB

Paper abstract: "We release code, models, and training data". "Hugging Face [01/22/2026] via https://huggingface.co/collections/nvidia/cosmos-policy"

LicenseNVIDIA One-Way Noncommercial License (NSCLv1)A

Model weights: "This model is released under the NVIDIA One-Way Noncommercial License (NSCLv1)"; "Models are not for commercial use." The code repository carries a separate licence: "Apache License Version 2.0, January 2004".
The developer’s own claims
Learns new tasksneeds retrainingB

Developer wording: "we introduce Cosmos Policy, a simple approach for adapting a large pretrained video model (Cosmos-Predict2) into an effective robot policy through a single stage of post-training on the robot demonstration data collected on the target platform, with no architectural modifications."

Works across robotsnoA

Model card for the ALOHA checkpoint: "Trained exclusively for ALOHA 2 robot platform with two ViperX 300 S robot arms"

Yes only if the developer itself says the model runs on different robot models, for example robots from two manufacturers or two of one manufacturer’s models. Several units of the same robot do not count.

Notes

  • NVIDIA Research adapts a pretrained video model (Cosmos-Predict2, not Cosmos 3) into a robot policy in a single post-training step; the policy also predicts future states and the value of a state.
  • NVIDIA released the code and models on 22 January 2026; there are separate checkpoints for the LIBERO and RoboCasa simulations and for the ALOHA robot.

Sources

One line per source, not per figure: a source has one retrieval date. The letter beside a figure points to the line it came from.