LeRobot/Cosmos3-Edge
lerobot/Cosmos3-Edge
This release offers a LeRobot policy conversion of the NVIDIA Cosmos3-Edge base checkpoint. It serves as a starting point for robot policy post-training without task-specific fine-tuning. The model jointly predicts robot action chunks and future video latents from language and current camera observations, with optional current-state conditioning. Usage requires a target dataset to resolve frame rate, camera size, and action layout. The package supports both DROID post-training recipes and LIBERO default configurations, with options to preserve or reinitialize action modules.
This repository provides a LeRobot-compatible conversion of the NVIDIA Cosmos3-Edge base checkpoint for robot policy post-training, without task-specific fine-tuning or reported closed-loop success rates.
Source: LeRobot — Model repository updates · Read original article ↗
Article text · Original source · English
On this page
Cosmos3-Edge base for LeRobot
This repository contains the LeRobot policy conversion of the NVIDIA Cosmos3-Edge base checkpoint. It is a starting point for robot policy post-training. The conversion adds no task-specific fine-tuning.
The policy jointly predicts robot action chunks and future video latents from language and current camera observations, with optional current-state conditioning. It contains the model components needed by the LeRobot policy interface.
| Item | Value |
|---|---|
| Original model developer | NVIDIA |
| Original checkpoint | nvidia/Cosmos3-Edge |
| Exact source revision | 344d602b128d1bbdacb43b08d0a3626f46343e29 |
| LeRobot compatibility baseline | b8f060837ced0a4ce4b136ea65b4d56bc7894bed |
| Policy type | cosmos3_edge |
| Backbone | Nemotron 3 Dense VL MoT |
| Decoder layers / hidden width | 28 / 2048 |
| Stored scalar weights, including frozen components and VAE | 4,563,688,396 |
| Stored tensors | 1188, comprising 1184 BF16 and 4 FP32 |
The stored scalar count is not a count of trainable or active parameters. The package includes a Wan2.2 VAE, tokenizer files, and serialized LeRobot processors.
Loading and post-training
Use a LeRobot version containing the Cosmos3 policy types. Conversion and CPU loading checks used the exact commit above.
The saved observation and action features are empty. Frame rate and camera size are unresolved. A target dataset must provide these values before constructing the policy. Read the configuration with:
from lerobot.configs.policies import PreTrainedConfig
config = PreTrainedConfig.from_pretrained("lerobot/Cosmos3-Edge")
The training entry points select different configurations:
--policy.path=lerobot/Cosmos3-Edgepreserves the saved DROID post-training recipe, including domain 8, 32-step chunks, state conditioning, identity normalization, UniPC with 4 inference steps, and guidance 3. State and action must already use the same layout and width.--policy.type=cosmos3_edge --policy.pretrained_path=lerobot/Cosmos3-Edgeuses these weights with the size class's default LIBERO recipe, including domain 5, 16-step chunks, no state conditioning, action quantile normalization, UniPC with 30 steps, and guidance 1. The dataset must match that recipe's camera and action layout.
load_action_modules defaults to true, preserving the released action modules. For native-style post-training, --policy.load_action_modules=false reinitializes the action input projection, action output projection, and action modality embedding. The setting returns to true after loading so subsequent reloads preserve trained action modules.
The saved dtype=float32 supports FP32 master weights during training. Most stored weights retain upstream BF16 precision. The four time-embedder tensors are FP32. For inference set --policy.dtype=bfloat16. For evaluation also set --tf32=false.
Scope and evaluation
The LeRobot implementation supports joint video-and-action wam mode, a single observation step, and the UniPC sampler. Actual action width comes from the target dataset. This package does not expose the upstream reasoning, audio, or standalone video-generation APIs.
CPU checks verified conversion and strict loading with resolved dataset features. No LIBERO closed-loop success rate is reported for this base conversion. Results for separately fine-tuned LIBERO checkpoints or NVIDIA Policy-DROID models do not measure this checkpoint.
License and provenance
The NVIDIA model materials use OpenMDW 1.1. Retain the license and applicable copyright and origin notices when redistributing them. NOTICE records the original model and conversion attribution. Separately installed runtime software retains its own licenses.
Consult the original NVIDIA model card and its safety, bias, privacy, and explainability notes for upstream limitations.
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
inference steps (UniPC)
4
View original evidence
pe, including domain 8, 32-step chunks, state conditioning, identity normalization, UniPC with 4 inference steps, and guidance 3. State and action must already use the same layout and width. --policy.type=cosmos3_edge --policy.pretrained_path=lerobot/Cosmos3-Edge uses these weights with the size class's default LIBERO recipe, including domain 5, 16-step chunks, no state conditioning, action quantile normalization, Un
Open source S6inference steps (UniPC)
30
View original evidence
iPC with 30 steps, and guidance 1. The dataset must match that recipe's camera and action layout. load_action_modules defaults to true , preserving the released action modules. For native-style post-training, --policy.load_action_modules=false reinitializes the action input projection, action output projection, and action modality embedding. The setting returns to true after loading so subsequent reloads preserve tra
Open source S7guidance value
3
View original evidence
pe, including domain 8, 32-step chunks, state conditioning, identity normalization, UniPC with 4 inference steps, and guidance 3. State and action must already use the same layout and width. --policy.type=cosmos3_edge --policy.pretrained_path=lerobot/Cosmos3-Edge uses these weights with the size class's default LIBERO recipe, including domain 5, 16-step chunks, no state conditioning, action quantile normalization, Un
Open source S6guidance value
1
View original evidence
iPC with 30 steps, and guidance 1. The dataset must match that recipe's camera and action layout. load_action_modules defaults to true , preserving the released action modules. For native-style post-training, --policy.load_action_modules=false reinitializes the action input projection, action output projection, and action modality embedding. The setting returns to true after loading so subsequent reloads preserve tra
Open source S7
Dataset details reported by the source
- License
- OpenMDW 1.1 Open source S10
- code: not_reported
- paper: not_reported
- dataset: not_reported
What remains unknown
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
venance The NVIDIA model materials use OpenMDW 1.1 . Retain the license and applicable copyright and origin notices when redistributing them. NOTICE records the original model and conversion attribution. Separately installed runtime software retains its own licenses. Consult the original NVIDIA model card and its safety , bias , privacy , and explainability notes for upstream limitations.
Open source S10
mponents needed by the LeRobot policy interface. Item Value Original model developer NVIDIA Original checkpoint nvidia/Cosmos3-Edge Exact source revision 344d602b128d1bbdacb43b08d0a3626f46343e29 LeRobot compatibility baseline b8f060837ced0a4ce4b136ea65b4d56bc7894bed Policy type cosmos3_edge Backbone Nemotron 3 Dense VL MoT Decoder layers / hidden width 28 / 2048 Stored scalar weights, including frozen components and
Open source S3
pe, including domain 8, 32-step chunks, state conditioning, identity normalization, UniPC with 4 inference steps, and guidance 3. State and action must already use the same layout and width. --policy.type=cosmos3_edge --policy.pretrained_path=lerobot/Cosmos3-Edge uses these weights with the size class's default LIBERO recipe, including domain 5, 16-step chunks, no state conditioning, action quantile normalization, Un
Open source S6
iPC with 30 steps, and guidance 1. The dataset must match that recipe's camera and action layout. load_action_modules defaults to true , preserving the released action modules. For native-style post-training, --policy.load_action_modules=false reinitializes the action input projection, action output projection, and action modality embedding. The setting returns to true after loading so subsequent reloads preserve tra
Open source S7
s are empty. Frame rate and camera size are unresolved. A target dataset must provide these values before constructing the policy. Read the configuration with: from lerobot.configs.policies import PreTrainedConfig config = PreTrainedConfig.from_pretrained( "lerobot/Cosmos3-Edge" ) The training entry points select different configurations: --policy.path=lerobot/Cosmos3-Edge preserves the saved DROID post-training reci
Open source S5
Implications for data suppliers
RoboSignal interpretation and collection questions, not statements of buyer demand.
- Confirm the required data type and collection setting with the buyer; this source does not establish a complete collection specification.
- Compare the reported units and scope before using these quantities in a budget. Recording hours, sensor-hours and trajectories are different measures.
- Check the exact dataset version, license, modalities and task coverage against the intended use.
- Validate demand and acceptance criteria with a buyer before scaling. Publication, popularity and a research result do not establish a purchase commitment.
Source：LeRobot — Model repository updates · huggingface.co