RoboCasa365 is a kitchen benchmark for a mobile
manipulator. ORA runs three of its target-split tasks on the PandaOmron
(a Panda arm on a holonomic base with a lifting torso): OpenDrawer,
PickPlaceCounterToCabinet and NavigateKitchen. The robot drives its base
with move_base, its arm follows admitted joint paths, and RoboCasa's own
_check_success decides success.
| Machine | x86_64 Linux; the simulator renders on the CPU (Mesa EGL), the perception services need an NVIDIA GPU |
| Disk | about 10 GB of kitchen assets plus the interpreter |
| Install | the explicit steps on RoboCasa365: install and run one episode |
| Run | backends.robocasa.spec compose, then the shared review and launch commands |
The one-command path (ora setup --env, ora run --env) does not cover
RoboCasa365 yet. The explicit path has four steps:
- Build the RoboCasa interpreter at the pinned robosuite and robocasa
revisions, download the kitchen assets and set
ROBOCASA_PYTHON. - Register a Mesa EGL slot (
runtime_kind: robocasa) in your host profile and qualify the perception services (shared setup). - Compose one cell, for example
OpenDraweron seed 0, with./scripts/ora spec robocasa compose. - Review and launch it; the result carries RoboCasa's verdict.
What to expect
Each seed samples its own kitchen and target, so the instruction comes from
that episode. The policy sees the task's horizon as its native-action
allowance (750 control steps for OpenDrawer and PickPlaceCounterToCabinet,
450 for NavigateKitchen), and every base or arm tick spends one. Arm targets
the robot cannot reach or that would run into the kitchen are refused before
any motion and cost nothing; the refusal says why.
More
- RoboCasa365 tasks: horizons and completion contracts.
- PandaOmron: the robot manual the model reads.
- RoboCasa365 and RoboLab backends: motion, admission, budget and qualification.