LIBERO is a set of Panda-arm kitchen and table tasks; LIBERO-Pro adds source-paired robustness variants of the same tasks (swapped objects or positions, paraphrased instructions, changed tasks). ORA drives the robot through perception, grasp-planning and motion tools; LIBERO's own termination check decides success.
There are two ways to run it:
| One NVIDIA GPU | Computer without a GPU | |
|---|---|---|
| Machine | x86_64 Linux, NVIDIA GPU (24 GB is enough for LIBERO) | macOS on Apple Silicon or Linux |
| Tools the model gets | full set: segmentation, grasp proposals, CuRobo motion planning | reduced set: bounded moves, gripper, RGB-D, point marking |
| Install | ./scripts/ora setup --env libero |
./scripts/ora setup |
| Run | ./scripts/ora run --env libero --task libero_goal/0 --confirm-launch |
./scripts/ora local --confirm-launch |
| Use it for | benchmark-comparable runs | trying ORA on your own computer |
Both need ORA itself (git clone -b preview/v0.1 https://github.com/fast-rua/ORA.git && cd ORA && uv sync --all-extras --locked) and a DeepSeek API key.
On one GPU
./scripts/ora setup --env libero
./scripts/ora providers key DEEPSEEK_LOW_API_KEY # once; or export DEEPSEEK_LOW_API_KEY=sk-...
./scripts/ora run --env libero --task libero_goal/0 --confirm-launch
- Install (~35 GB, mostly downloads; CuRobo builds in about 6 minutes): LIBERO-Pro sources and task data at pinned revisions, a Python 3.11 simulator environment with CUDA 12.8 PyTorch, CuRobo built for your GPU, the SAM3 and Contact-GraspNet perception services, the agent core's Pi packages and the terminal UI. Check the tools to install first.
- Run.
ora runstarts the perception services if needed (about 3 minutes the first time), writes the run's specification and opens the terminal UI. Press Enter on the task in the picker. LIBERO needs no one-time qualification. - Result on an RTX PRO 6000:
libero_goal/0("open the middle drawer of the cabinet") reached the official success in 5 decisions and 11 model calls; the episode took about 4 minutes.
Choosing a task
--task <suite>/<id>, with --seed 0..49 for the initial state (default 0):
| Suites | Tasks |
|---|---|
libero_spatial, libero_object, libero_goal, libero_10 |
ids 0–9 in each |
LIBERO-Pro: <suite>_task, <suite>_swap, <suite>_lan, <suite>_object (for example libero_goal_swap) |
the same ids, with the perturbation applied |
./scripts/ora run --env libero --task libero_spatial/3 --seed 7 --confirm-launch
./scripts/ora run --env libero --task libero_10_lan/2 --confirm-launch
On a computer without a GPU
./scripts/ora setup # ~2 GB: LIBERO sources, a CPU simulator environment, the agent core, the terminal UI
./scripts/ora providers key DEEPSEEK_LOW_API_KEY # once; or export DEEPSEEK_LOW_API_KEY=sk-...
./scripts/ora doctor # "ready for task selection"
./scripts/ora local --confirm-launch # pick a suite and task in the picker
This uses a reduced tool set, so its results are not comparable with the GPU
setup. On Linux, libero_goal/0 reached the official success after 40 actions
and 70 model calls in about 21 minutes. Details: Run ORA on your computer.
While it runs
Enter sends the agent a message (the run is then marked human-assisted), Esc
stops it, Ctrl-C leaves the UI with the run still going. Replay any run with
./scripts/ora watch RUN_ID. See Watch, steer and inspect.
More
- One-GPU guide: requirements, GPU and driver notes, troubleshooting.
- Explicit LIBERO installation: each pinned source and package, for a custom or multi-GPU host.