Added poetry setup

[aliberts]

Added poetry setup

[chaofiber]

Review

Positives

1. Huge DX improvement — The old setup required cloning 3 repos, running setup.py develop on each, manually pip install-ing ~12 packages, and a separate Hydra fix. Poetry reduces this to poetry install. Meaningful reduction in onboarding friction.

2. Reproducible builds — The lockfile pins exact versions, which the previous workflow did not.

3. Clean README rewrite — Concise instructions with a useful troubleshooting tip for disk space issues.

---

Issues & Concerns

1. Dual diffusion_policy installation paths

The README instructs users to:

git clone https://github.com/real-stanford/diffusion_policy
cp -r diffusion_policy/diffusion_policy $(poetry env info -p)/lib/python3.10/site-packages/

But pyproject.toml also declares diffusion-policy as a git dependency. Two conflicting mechanisms for the same package — confusing and fragile. The cp hack is unversioned, invisible to the package manager, and hardcodes python3.10. Pick one approach and remove the other.

2. Dependency issues in pyproject.toml

• tensordict pins to HEAD of main (will drift silently), while torchrl pins a specific commit (13bef426). Both should pin commits or track tagged releases for reproducibility.
• mujoco-py = "^2.1.2.14" + mujoco = "^3.1.2": mujoco-py is the old deprecated binding requiring a MuJoCo license; mujoco is the new official package. Having both is contradictory and can cause conflicting libc requirements.
• gym = "^0.26.2": By 2024, gymnasium had replaced gym as the maintained fork. gym is a source of deprecation warnings and API incompatibilities with modern RL tooling.
• No optional/extra groups: All deps (MuJoCo, pygame, GPU-specific torch) are mandatory. Poetry supports dependency groups — users who only need dataset loading shouldn't be forced to install heavy simulation deps.

3. Python classifier is wrong

classifiers=[
    "Programming Language :: Python :: 3.8",
]

But python = "^3.10". The classifier claims 3.8 support that doesn't exist.

4. Minor: 3 trailing blank lines before [build-system] in pyproject.toml

---

Overall the core decision to adopt a package manager was clearly correct and this was a good stepping stone for the project.