NVIDIA Releases Cosmos3-Nano Policy Model Trained on LIBERO-10
nvidia/cosmos3-nano-libero10-sdg-export-S2_s43
NVIDIA has released a trained policy model for the Cosmos3-Nano robot, post-trained on the LIBERO-10 environment. The model, based on training seed 43, achieves 94.9% success in standard scenes and 89.7% in perturbed scenes. The model is part of a synthetic-data study and is available via the cosmos-framework policy server.
NVIDIA has released a trained policy model for the Cosmos3-Nano robot, post-trained on the LIBERO-10 environment. The model, based on training seed 43, achieves 94.9% success in standard scenes and 89.7% in perturbed scenes. The model is part of a synthetic-data study and is available via the cosmos-framework policy server.
Source: NVIDIA — Robotics model repository updates · Read original article ↗
Article text · Original source · English
Cosmos3-Nano LIBERO-10 policy, cell S2_s43 (training seed 43)
A Cosmos3-Nano action policy post-trained on LIBERO-10, the second training seed of 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 + 1137 restyled from iteration 1000 (training seed 43) |
| Iterations | resume R(seed 43) at 1000 -> 2000 (global batch 2048, training seed 43) |
| Success, standard scenes | 94.9 +/- 0.4 % (10 tasks x 50 initial states, three evaluation seeds; values reported by the training run, record pending) |
| Success, perturbed scenes | 89.7 +/- 1.8 % (same initial states with randomised lighting, textures, skybox and camera jitter; measured on the evaluation bed, 2026-10-19) |
| Control | D2_s43 (4.2 points on perturbed scenes) |
| 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
43
View original evidence
Cosmos3-Nano LIBERO-10 policy, cell S2_s43 (training seed 43) A Cosmos3-Nano action policy post-trained on LIBERO-10, the second training seed of 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
Open source S2training batch size
2,048
View original evidence
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 + 1137 restyled from iteration 1000 (training seed 43) Iterations resume R(seed 43) at 1000 -> 2000 (global batch 2048, training seed 43) Success, standard scenes 94.9 +/- 0.4
Open source S3training iterations
>2,000
View original evidence
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 + 1137 restyled from iteration 1000 (training seed 43) Iterations resume R(seed 43) at 1000 -> 2000 (global batch 2048, training seed 43) Success, standard scenes 94.9 +/- 0.4
Open source S3training data size
379
View original evidence
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 + 1137 restyled from iteration 1000 (training seed 43) Iterations resume R(seed 43) at 1000 -> 2000 (global batch 2048, training seed 43) Success, standard scenes 94.9 +/- 0.4
Open source S3restyled data size
1,137
View original evidence
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 + 1137 restyled from iteration 1000 (training seed 43) Iterations resume R(seed 43) at 1000 -> 2000 (global batch 2048, training seed 43) Success, standard scenes 94.9 +/- 0.4
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.
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 + 1137 restyled from iteration 1000 (training seed 43) Iterations resume R(seed 43) at 1000 -> 2000 (global batch 2048, training seed 43) Success, standard scenes 94.9 +/- 0.4
Open source S3
% (10 tasks x 50 initial states, three evaluation seeds; values reported by the training run, record pending) Success, perturbed scenes 89.7 +/- 1.8 % (same initial states with randomised lighting, textures, skybox and camera jitter; measured on the evaluation bed, 2026-10-19) Control D2_s43 (4.2 points on perturbed scenes) Base model nvidia/Cosmos3-Nano , revision e59a53c2 Framework cosmos-framework c23e51f2 (v1.2.
Open source S4
Cosmos3-Nano LIBERO-10 policy, cell S2_s43 (training seed 43) A Cosmos3-Nano action policy post-trained on LIBERO-10, the second training seed of 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
Open source S2
Source：NVIDIA — Robotics model repository updates · huggingface.co