Files
UniversalHumanoid/.agents/skills/setup-env/SKILL.md
T
2026-11-03 19:46:08 +08:00

2.1 KiB

name, description
name description
setup-env Detect the host, install the locked vendored runtime in a fresh environment, and verify dependencies, backend origins, CUDA and robot gates.

Set up and verify an environment

Use for first installation, another machine, missing imports, wrong backend, or CUDA failures. Read Setup and Vendor provenance.

Procedure

  1. Run python3 .agents/skills/setup-env/scripts/detect.py. Record Python/platform, GPU/driver, installer availability and existing import origins. Detection is not a simulation pass.
  2. Confirm Linux x86-64 and compatible NVIDIA CUDA support before choosing the reference lock. Do not substitute CPU training or silently change Torch/Warp versions.
  3. Run uv sync --locked --extra dev --python 3.13 in a fresh project environment. The ordinary pip install -e . path and hashed pip alternative are documented in Setup. Never install separate mjlab/rsl-rl-lib copies into this environment or repair another project's venv.
  4. Run ./scripts/uh doctor, ./scripts/uh verify-vendor, and ./scripts/check.
  5. Run .venv/bin/python .agents/skills/setup-env/scripts/check.py --config experiments/g1_smoke.json --output runs/setup-g1. Use the T1 smoke config for T1. Read every gate log and summary.json, not just a GPU name.

Failure handling

  • A source mismatch means wrong installation or an unreviewed vendor change. Inspect it; do not refresh hashes or unset errors to manufacture success.
  • A locked wheel download can be slow. Report installation as pending until the installer exits successfully and dependency checks pass.
  • A robot check failure belongs to the new-task procedure; changing rewards cannot repair an IMU frame.
  • An unavailable GPU/driver is an external prerequisite, not permission to claim CPU equivalence.

Handoff

Report interpreter, installed backend origins, lock/inventory status, GPU/driver, gate output paths and remaining failures. Only then authorize a bounded training smoke within the operator's budget. No credential collection, remote installation or hardware operations are part of setup.