Found during internal QA of the 1.4 release.
The Televiz Quick Start tells you to install with pip install isaacteleop, and then the first runnable snippet on the same page opens with import cupy as cp. cupy is not a dependency of the wheel and is never listed as a prerequisite, so the snippet fails on a clean environment with ModuleNotFoundError: No module named 'cupy'. It runs once cupy is installed by hand.
Where
Why it fails
cupy appears in no dependency list in pyproject.toml — not in [project] dependencies (numpy, pyyaml) and not in any extra (dev, ui, retargeters, retargeters-lite, grounding, cloudxr). Verified identical on main and release/1.4.x.
Suggested fix
Either list cupy as a prerequisite ahead of the snippet — with the CUDA-matched package name, e.g. cupy-cuda12x, since a bare pip install cupy is not what most people want — or rewrite the snippet against a VizBuffer / NumPy path so the Quick Start runs on a bare pip install isaacteleop. The page already notes that any __cuda_array_interface__ producer works (CuPy / PyTorch / Numba), so CuPy is illustrative here rather than required, which argues for the second option.
Found during internal QA of the 1.4 release.
The Televiz Quick Start tells you to install with
pip install isaacteleop, and then the first runnable snippet on the same page opens withimport cupy as cp.cupyis not a dependency of the wheel and is never listed as a prerequisite, so the snippet fails on a clean environment withModuleNotFoundError: No module named 'cupy'. It runs oncecupyis installed by hand.Where
docs/source/getting_started/televiz.rstL32:pip install isaacteleopdocs/source/getting_started/televiz.rstL103:import cupy as cpWhy it fails
cupyappears in no dependency list inpyproject.toml— not in[project] dependencies(numpy,pyyaml) and not in any extra (dev,ui,retargeters,retargeters-lite,grounding,cloudxr). Verified identical onmainandrelease/1.4.x.Suggested fix
Either list
cupyas a prerequisite ahead of the snippet — with the CUDA-matched package name, e.g.cupy-cuda12x, since a barepip install cupyis not what most people want — or rewrite the snippet against aVizBuffer/ NumPy path so the Quick Start runs on a barepip install isaacteleop. The page already notes that any__cuda_array_interface__producer works (CuPy / PyTorch / Numba), so CuPy is illustrative here rather than required, which argues for the second option.