NVIDIA publishes Cosmos3-Nano LIBERO-10 robot policy checkpoints on Hugging Face

AINVIDIA releases a Cosmos3-Nano action policy post-trained on LIBERO-10 with training seed 43, as a bf16 checkpoint of about 30 GB. The model scores 91.8 ± 0.5% on standard LIBERO-10 scenes and 86.8 ± 1.1% on perturbed scenes, which NVIDIA measured on its evaluation bed. The training data is not redistributed, and the checkpoint is governed by the NVIDIA Open Model License.

AINVIDIA publishes a Cosmos3-Nano action policy post-trained on LIBERO-10 with training seed 43, as a bf16 checkpoint of about 30 GB. The policy scores 93.3 ± 0.7% on standard LIBERO-10 scenes and 85.5 ± 2.5% on perturbed scenes with randomized lighting, textures, skybox and camera jitter. The training data is not redistributed; LIBERO-10 is available at nvidia/LIBERO_LeRobot_v3, and the checkpoint is governed by the NVIDIA Open Model License.

AINVIDIA publishes a Cosmos3-Nano action policy post-trained on LIBERO-10, a bf16 safetensors checkpoint of about 30 GB loadable by cosmos-framework. The policy scores 94.9 ± 0.4% success on standard scenes and 89.7 ± 1.8% on perturbed scenes, which the source reports against a control model at 4.2 points lower on perturbed scenes. Training data is not redistributed, and differences under 2.5 points fall within training-seed spread, according to the source.