How is pricing calculated?

Pricing is based on the hourly rate of operators in your selected region. You only pay for the time operators spend controlling your robots.

What robots are supported?

We support 9+ robot arms including SO-100, Koch, Franka FR3, Franka Panda, WidowX 250, ViperX 300, ViperX 300s, and ALOHA.

What is the latency for teleoperation?

Our platform provides ultra-low latency of 30-70ms worldwide, enabling real-time responsive robot control.

Cosmos Policy

NVIDIA, United States · January 2026

NVIDIA and Stanford University

World modelOpen weights, Code Apache-2.0Also written Cosmos-Policy-Predict2-2B, NVIDIA Cosmos Policy

Parameters: 2 B; 2B parameters.
GPU memory: 6.8 GB; reported by the authors
Inference latency: not published; per action step
Weights: Code Apache-2.0; Code Apache-2.0.

What it is

Cosmos Policy takes the opposite route to every VLA in this comparison: instead of bolting an action head onto a vision language model, it post-trains NVIDIA's Cosmos-Predict2-2B video world model into a robot policy. Actions, proprioception and value estimates are injected as latent frames inside the video diffusion sequence, so no architectural change is needed and the model keeps its ability to predict future states, which enables test-time best-of-N planning. It posts the best simulation numbers here, 98.5 percent average on LIBERO and 67.1 percent on RoboCasa, and 93.6 average points on real bimanual ALOHA tasks. Two practical caveats decide whether it fits a project: planning mode needs about 5 seconds per action chunk, and the weights carry a non-commercial NVIDIA licence even though the code is Apache-2.0.

Architecture

Backbone: Cosmos-Predict2-2B-Video2World latent video diffusion model, using the Wan2.1 spatiotemporal VAE tokenizer and a diffusion transformer. Newer checkpoints built on Cosmos-Predict2.5-2B exist for LIBERO and RoboCasa.
Action head: Latent frame injection. Actions, proprioception and value predictions are encoded as latent frames inside the video diffusion sequence with no architectural modification, decoded with 5 denoising steps. Action chunk length is task specific: 16 timesteps on LIBERO, 32 on RoboCasa with 16 executed before requery, 50 on ALOHA.
Action chunk: 16 steps per forward pass
Control rate: 25 Hz
Parameters: 2B parameters. Released as task-specific checkpoints rather than one generalist model: Cosmos-Policy-LIBERO-Predict2-2B, Cosmos-Policy-RoboCasa-Predict2-2B, Cosmos-Policy-ALOHA-Predict2-2B and a separate ALOHA planning model for world-model and value predictions. Predict2.5-based variants for LIBERO and RoboCasa were added in February 2026.
Pretraining data: Inherited from the Cosmos-Predict2 video world foundation model, which is pretrained on large-scale video. Cosmos Policy itself adds a single stage of post-training on robot demonstrations collected on the target platform. RoboCasa results use 50 demonstrations per task.
Embodiments: Franka Emika Panda (LIBERO simulation), RoboCasa kitchen simulation, ALOHA bimanual real robot

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Computed from 2 B parameters. Weights only, so treat it as a floor: activations, image encoder intermediates and the CUDA context come on top, in practice 30 to 50 percent more for inference.
Precision	Weights	Note
fp32	7.5 GB	Training master weights, rarely used for inference
bf16 or fp16	3.7 GB	The usual way these checkpoints are served
int8	1.9 GB	Quantised, expect some loss of precision on fine motions
int4	0.9 GB	Aggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 8 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.

Published figures. These come from the model authors, not from this site.
Inference memory, reported	6.8 GB source
Control rate	25 Hz

The model card lists inference only: 6.8 GB for the LIBERO tasks, 8.9 GB for RoboCasa and 6.0 GB for the ALOHA tasks. Training ran on a cluster of 64 H100 cards, which is not a per GPU requirement.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

Reproducible benchmark suites. Everyone runs the same episodes, so the numbers can be compared inside a suite.

LIBERO average over Spatial, Object, Goal and Long, 500 trials per suite, 3 seeds
98.5%
success rateAuthors' own evaluation. Per suite: Spatial 98.1, Object 100.0, Goal 98.2, Long 97.6. The HF model card reports 6,000 trials in total.source
LIBERO Long
97.6%
success rateAuthors' own evaluation. The hardest LIBERO suite and the more discriminative of the four.source
RoboCasa 24 kitchen manipulation tasks, 3,600 trials, 50 demonstrations per task
67.1%
success rateAuthors' own evaluation. Far below the near-saturated LIBERO number, which makes it the more useful comparison point.source

Real world

Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.

Real-world bimanual ALOHA average over 4 tasks, 101 trials, 25 Hz control
93.6 points
average scoreAuthors' own evaluation, partial-credit scoring rather than binary success. Per task: put X on plate 100.0, fold shirt 99.5, put candies in bowl 89.6, put candy in ziploc bag 85.4.source
Real-world bimanual ALOHA put candy in ziploc bag
85.4 points
scoreAuthors' own evaluation. Weakest of the four real tasks.source

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Real-world validation is on an ALOHA bimanual setup at 25 Hz with 50-timestep (2 second) action chunks, 101 trials in total. Reported scores: put X on plate 100.0, fold shirt 99.5, put candies in bowl 89.6, put candy in ziploc bag 85.4, average 93.6 points.

Fine tuning it yourself

A single stage of post-training on demonstrations from the target platform, with no architectural modifications. Inputs are multi-view RGB at 224x224, a text task description and a 9-dimensional proprioceptive state. Outputs are a 16-timestep action sequence plus future state predictions and value estimates. RoboCasa results were obtained with 50 demonstrations per task. Recipes are published in the Cosmos Cookbook, and training data is released alongside the checkpoints. Effective planning additionally requires substantial policy rollout data beyond the demonstration distribution.

Where it helps, where it does not

Strengths

Only model here that starts from a video world foundation model instead of a VLM, which brings learned temporal causality and implicit physics from video pretraining.
Retains future-state prediction and value estimation, enabling test-time best-of-N planning that a plain policy cannot do.
Lowest documented inference footprint in this comparison: one GPU with 6.0 to 8.9 GB VRAM, which fits consumer hardware.
Best published simulation numbers here: 98.5 percent on LIBERO and 67.1 percent on RoboCasa, plus 93.6 average points on real ALOHA tasks.
Checkpoints, training data and post-training recipes are all published, including newer Predict2.5-based variants.
Post-training needs no architectural modification, just one stage on target-platform demonstrations.

Limits

Weights are under the NVIDIA One-Way Noncommercial License (NSCLv1) and are non-commercial only, even though the code is Apache-2.0. This rules out product use.
Model-based planning is slow, around 5 seconds per action chunk according to the authors, which they say may limit applicability to dynamic tasks. The measurement hardware is not named.
Released as task-specific checkpoints (LIBERO, RoboCasa, ALOHA), not a generalist cross-embodiment policy. The LIBERO checkpoint is trained exclusively for a Franka Emika Panda in simulation and needs a specific dual-camera setup.
Effective planning requires substantial policy rollout data beyond the demonstration distribution, not just demonstrations.
Training is expensive: 8 to 64 H100 80 GB GPUs for about 48 hours.
Planning uses best-of-N search with a single layer in the search tree, so it is shallow compared to real tree search.
LIBERO at 98.5 percent is near saturation and carries little discriminative signal. RoboCasa at 67.1 percent is the honest measure of remaining headroom.

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-08-11.

Try a policy on a real arm

A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.

Download the client Drive a real arm now