NVIDIA Releases Cosmos3-Nano Policy Model Trained on LIBERO-10
nvidia/cosmos3-nano-libero10-sdg-export-S_best
NVIDIA has released a trained policy model for the Cosmos3-Nano robot, fine-tuned on the LIBERO-10 environment. The model achieves 93.3% success in standard scenes and 83.2% in perturbed scenes. It is part of a synthetic data study and is licensed under NVIDIA's Open Model License.
NVIDIA has released a trained policy model for the Cosmos3-Nano robot, specifically fine-tuned on the LIBERO-10 environment. The model demonstrates success rates in both standard and perturbed scenes, with performance metrics provided. The model is part of a broader study involving synthetic data restyling and is licensed under NVIDIA's Open Model License.
Source: NVIDIA — Robotics model repository updates · Read original article ↗
Article text · Original source · English
Cosmos3-Nano LIBERO-10 policy, cell S_best
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 + 568 restyled the gate scored best |
| Iterations | 0->2000 (global batch 2048, training seed 42) |
| Success, standard scenes | 93.3 +/- 1.2 % (10 tasks x 50 initial states, three evaluation seeds) |
| Success, perturbed scenes | 83.2 +/- 1.6 % (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 iterations
>2,000
View original evidence
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 + 568 restyled the gate scored best Iterations 0->2000 (global batch 2048, training seed 42) Success, standard scenes 93.3 +/- 1.2 % (10 tasks x 50 initial states, three evaluation seeds) Success, perturbed scenes 83.2 +/-
Open source S3training seed
42
View original evidence
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 + 568 restyled the gate scored best Iterations 0->2000 (global batch 2048, training seed 42) Success, standard scenes 93.3 +/- 1.2 % (10 tasks x 50 initial states, three evaluation seeds) Success, perturbed scenes 83.2 +/-
Open source S3evaluation seeds
3
View original evidence
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 + 568 restyled the gate scored best Iterations 0->2000 (global batch 2048, training seed 42) Success, standard scenes 93.3 +/- 1.2 % (10 tasks x 50 initial states, three evaluation seeds) Success, perturbed scenes 83.2 +/-
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.
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 + 568 restyled the gate scored best Iterations 0->2000 (global batch 2048, training seed 42) Success, standard scenes 93.3 +/- 1.2 % (10 tasks x 50 initial states, three evaluation seeds) Success, perturbed scenes 83.2 +/-
Open source S3
Source：NVIDIA — Robotics model repository updates · huggingface.co