# RX1 v2 — learned action model + Qwen2.5-VL planner.
# The control pipeline (collect/train/run the learned policy) needs only the
# core stack; the VLM planner is optional and GPU-oriented.

# ── core (CPU is enough to collect demos, train, and run the learned policy) ──
mujoco>=3.1.0
numpy>=1.24.0
torch>=2.1.0
PyYAML>=6.0
Pillow>=10.0.0

# ── Qwen2.5-VL-3B high-level planner (optional; needs an NVIDIA GPU) ──────────
# 4-bit loading fits the 3B model on ~6 GB VRAM.
transformers>=4.49.0
accelerate>=0.30.0
bitsandbytes>=0.43.0        # 4-bit quantisation (CUDA only)
qwen-vl-utils>=0.0.8

# For a real Cosmos world model (heavy; not needed for the learned controller):
#   pip install -e cosmos-framework      # or  diffusers>=0.33