NVIDIA Releases Cosmos3-Nano Policy Model Trained on LIBERO-10

nvidia/cosmos3-nano-libero10-sdg-export-D5

Summary

NVIDIA has released a robotics policy model, Cosmos3-Nano LIBERO-10, trained on the LIBERO-10 environment. The model demonstrates success rates of 93.4% in standard scenes and 82.9% in perturbed scenes, with performance evaluated across 10 tasks and 50 initial states. The model is available under the NVIDIA Open Model License.

Editorial context

NVIDIA provides an open-source robotics policy model trained on LIBERO-10, with performance metrics reported across standard and perturbed scenes, highlighting the model's adaptability to varied conditions.

Source: NVIDIA — Robotics model repository updates · Read original article ↗

Article text · Original source · English

Cosmos3-Nano LIBERO-10 policy, cell D5

A Cosmos3-Nano action policy post-trained on LIBERO-10, one cell of the controlled synthetic-data study in the Cosmos Cookbook recipe "Restyle LIBERO-10 Demonstrations with Cosmos3-Nano-Sim-Transfer and Measure the Gain Against a Matched Control". Export of the training checkpoint (bf16 safetensors, EMA weights, 7 shards, about 30 GB); loadable by the cosmos-framework policy server. The training data is not redistributed here: LIBERO-10 is at nvidia/LIBERO_LeRobot_v3, and the restyled sets are regenerated by the recipe.


Training data	379 real LIBERO-10 demonstrations + 1895 duplicates
Iterations	0->2000 (global batch 2048, training seed 42)
Success, standard scenes	93.4 +/- 1.2 % (10 tasks x 50 initial states, three evaluation seeds)
Success, perturbed scenes	82.9 +/- 1.9 % (same initial states with randomised lighting, textures, skybox and camera jitter)
Base model	`nvidia/Cosmos3-Nano`, revision `e59a53c2`
Framework	`cosmos-framework` `c23e51f2` (v1.2.2)

Read the recipe before comparing cells: a restyled cell is only interpretable next to its duplicate control, and differences under 2.5 points are within the training-seed spread.

Licence: derivative of nvidia/Cosmos3-Nano; the NVIDIA Open Model License applies.

What the source reports

Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.

Reported numbers

training seed
42
View original evidence
ver. The training data is not redistributed here: LIBERO-10 is at nvidia/LIBERO_LeRobot_v3 , and the restyled sets are regenerated by the recipe. Training data 379 real LIBERO-10 demonstrations + 1895 duplicates Iterations 0->2000 (global batch 2048, training seed 42) Success, standard scenes 93.4 +/- 1.2 % (10 tasks x 50 initial states, three evaluation seeds) Success, perturbed scenes 82.9 +/- 1.9 % (same initial s
Open source S3
evaluation seeds
3
View original evidence
ver. The training data is not redistributed here: LIBERO-10 is at nvidia/LIBERO_LeRobot_v3 , and the restyled sets are regenerated by the recipe. Training data 379 real LIBERO-10 demonstrations + 1895 duplicates Iterations 0->2000 (global batch 2048, training seed 42) Success, standard scenes 93.4 +/- 1.2 % (10 tasks x 50 initial states, three evaluation seeds) Success, perturbed scenes 82.9 +/- 1.9 % (same initial s
Open source S3
initial states
50
View original evidence
ver. The training data is not redistributed here: LIBERO-10 is at nvidia/LIBERO_LeRobot_v3 , and the restyled sets are regenerated by the recipe. Training data 379 real LIBERO-10 demonstrations + 1895 duplicates Iterations 0->2000 (global batch 2048, training seed 42) Success, standard scenes 93.4 +/- 1.2 % (10 tasks x 50 initial states, three evaluation seeds) Success, perturbed scenes 82.9 +/- 1.9 % (same initial s
Open source S3

code: not_reported
weights: not_reported
dataset: not_reported
paper: not_reported

Source excerpts and review record

Automatically extracted; no manual editorial approval recorded.

Cosmos3-Nano LIBERO-10 policy, cell D5 A Cosmos3-Nano action policy post-trained on LIBERO-10, one cell of the controlled synthetic-data study in the Cosmos Cookbook recipe "Restyle LIBERO-10 Demonstrations with Cosmos3-Nano-Sim-Transfer and Measure the Gain Against a Matched Control". Export of the training checkpoint (bf16 safetensors, EMA weights, 7 shards, about 30 GB); loadable by the cosmos-framework policy ser
Open source S2

ver. The training data is not redistributed here: LIBERO-10 is at nvidia/LIBERO_LeRobot_v3 , and the restyled sets are regenerated by the recipe. Training data 379 real LIBERO-10 demonstrations + 1895 duplicates Iterations 0->2000 (global batch 2048, training seed 42) Success, standard scenes 93.4 +/- 1.2 % (10 tasks x 50 initial states, three evaluation seeds) Success, perturbed scenes 82.9 +/- 1.9 % (same initial s
Open source S3