README.md

metadata

license: other
library_name: cosmos-policy
pipeline_tag: robotics
tags:
  - robotics
  - cosmos-policy
  - openarm
  - lerobot
  - action-generation

Cosmos Policy OpenArm Predict2 2B

Fine-tuned NVIDIA Cosmos Policy checkpoint for OpenArm using asterisms45/my_openarm_dataset_combined_2cam.

This repository contains the inference checkpoint only:

• model/: Cosmos distributed checkpoint model weights from openarm-2cam-46/checkpoints/iter_000050000/model
• openarm_dataset_statistics.json: action/proprio normalization statistics
• openarm_t5_embeddings.pkl: cached T5 task embeddings

The optimizer state from the training run is intentionally not uploaded.

Example

uv run --extra cu128 --group aloha --python 3.10 python -m \
  cosmos_policy.experiments.robot.aloha.deploy \
  --config cosmos_predict2_2b_480p_openarm_asterisms45_2cam_posttrain__inference_only \
  --ckpt_path asterisms45/Cosmos-Policy-OpenArm-Predict2-2B \
  --config_file cosmos_policy/config/config.py \
  --use_third_person_image True \
  --use_wrist_image True \
  --num_wrist_images 2 \
  --use_proprio True \
  --normalize_proprio True \
  --unnormalize_actions True \
  --dataset_stats_path asterisms45/Cosmos-Policy-OpenArm-Predict2-2B/openarm_dataset_statistics.json \
  --t5_text_embeddings_path asterisms45/Cosmos-Policy-OpenArm-Predict2-2B/openarm_t5_embeddings.pkl \
  --trained_with_image_aug True \
  --chunk_size 50 \
  --num_open_loop_steps 50 \
  --num_denoising_steps_action 10 \
  --deterministic True

Model weights derive from NVIDIA Cosmos Policy. Check the upstream repository and model license terms before redistribution or commercial use.