You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
RoboCasa currently has only CPU mock tests (tests/envs/test_robocasa.py): registration, factory dispatch, per-rank sharding, the version shim — none of which touch the real sim. LIBERO, by contrast, is exercised in gpu_test.yml and the nightly regression_test.yml, and has a config-wiring test (tests/envs/test_libero_control_freq.py).
So there's no sim-level regression protection for RoboCasa: a break in the wrapper / robosuite-master / robocasa-fork stack wouldn't be caught by CI.
Why it matters
The RoboCasa stack has several moving parts (robosuite pinned to a master commit, a packaging fork, a version shim, EGL rendering). Without GPU/regression coverage, regressions in any of those land silently.
Suggested approach
A gpu-marked smoke test (construct / reset / step / render on a couple of tasks), gated on RoboCasa asset availability (skip cleanly when assets aren't present, since they're a ~5-10 GB runtime download).
Decide how/whether to cache the assets for the GPU CI runners.
Problem
RoboCasa currently has only CPU mock tests (
tests/envs/test_robocasa.py): registration, factory dispatch, per-rank sharding, the version shim — none of which touch the real sim. LIBERO, by contrast, is exercised ingpu_test.ymland the nightlyregression_test.yml, and has a config-wiring test (tests/envs/test_libero_control_freq.py).So there's no sim-level regression protection for RoboCasa: a break in the wrapper / robosuite-master / robocasa-fork stack wouldn't be caught by CI.
Why it matters
The RoboCasa stack has several moving parts (robosuite pinned to a master commit, a packaging fork, a version shim, EGL rendering). Without GPU/regression coverage, regressions in any of those land silently.
Suggested approach
gpu-marked smoke test (construct / reset / step / render on a couple of tasks), gated on RoboCasa asset availability (skip cleanly when assets aren't present, since they're a ~5-10 GB runtime download).control_freqwiring unit test once RoboCasa: env.fps is inert — control_freq not wired to the sim #380 lands (configfps->gym_kwargs["control_freq"]-> sim), mirroringtest_libero_control_freq.py.References
tests/envs/test_robocasa.py,.github/workflows/gpu_test.yml,.github/workflows/regression_test.yml