mujoco>=3.1.0
numpy>=1.24.0
torch>=2.1.0
torchvision>=0.16.0
Pillow>=10.0.0
PyYAML>=6.0
opencv-python>=4.8.0
einops>=0.7.0
huggingface_hub>=0.21.0
transformers>=4.40.0
accelerate>=0.27.0
peft>=0.9.0
timm>=0.9.0

# ── Cosmos world model ──────────────────────────────────────────────────────
# Option A (recommended): cosmos-framework — the actual NVIDIA inference library
#   git clone https://github.com/NVIDIA/cosmos-framework
#   pip install -e cosmos-framework
#   huggingface-cli login  (accept gated license on HF for Cosmos-Predict2.5-2B)
#
# Option B: diffusers — lighter weight, covers most Cosmos-1.0 and Cosmos-Predict2 models
#   pip install "diffusers>=0.33"
#   pip install av imageio imageio-ffmpeg  # video I/O deps

# ── Cosmos Policy ───────────────────────────────────────────────────────────
# Clone + install (NOT on PyPI):
#   git clone https://github.com/NVlabs/cosmos-policy
#   pip install -e cosmos-policy
#   huggingface-cli login  (NVIDIA One-Way Noncommercial License)
#
# Available checkpoints:
#   nvidia/Cosmos-Policy-LIBERO-Predict2-2B     (7-DoF simulated arm — best for MuJoCo start)
#   nvidia/Cosmos-Policy-RoboCasa-Predict2-2B   (7-DoF simulated arm)
#   nvidia/Cosmos-Policy-ALOHA-Predict2-2B      (14-DoF bimanual real robot)
#
# !! Action space: all checkpoints output 7-14 DoF; humanoid needs 21 DoF.
#    Fine-tune the action head on humanoid demonstrations before real deployment.
#    Training scripts: https://github.com/NVlabs/cosmos-policy (see docs/training.md)

# ── OpenVLA language grounding ──────────────────────────────────────────────
# pip install git+https://github.com/openvla/openvla.git
# huggingface-cli login  (model: openvla/openvla-7b)