RoboTwin 2.0 is a dual-arm (ALOHA)
manipulation benchmark with 50 tasks in clean or randomized scenes. ORA drives
both arms through perception, grasp-planning and motion tools; RoboTwin's own
check_success decides success.
| Machine | x86_64 Linux with one NVIDIA GPU, 32 GB or more |
| Disk | ~57 GB (14 GB of asset archives, unpacked to 16 GB) |
| Install | ./scripts/ora setup --env robotwin |
| Run | ./scripts/ora run --env robotwin --task click_bell --confirm-launch |
You 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.
Install and run
./scripts/ora setup --env robotwin
./scripts/ora providers key DEEPSEEK_LOW_API_KEY # once; or export DEEPSEEK_LOW_API_KEY=sk-...
./scripts/ora run --env robotwin --task click_bell --confirm-launch
-
Install. Setup installs, mostly as downloads:
- RoboTwin at its pinned revision with ORA's vendored planner install;
- a Python 3.10 environment with CUDA 12.8 PyTorch and the upstream SAPIEN and MPLib;
- CuRobo built for your GPU (about 6 minutes);
- the RoboTwin 2.0 assets;
- the SAM3 and Contact-GraspNet services;
- the agent core's Pi packages and the terminal UI.
Check the tools to install first.
-
First run of a task. RoboTwin needs a seed that the upstream expert policy can solve on your machine.
ora runshows an estimate, asks, then scans the task's seeds and keeps the first that works: about 2 minutes forclick_bell. Later runs of that task reuse the record. -
Run. The perception services start if needed (about 3 minutes the first time). Then the terminal UI opens with the task; press Enter in the picker.
-
Result on an RTX PRO 6000:
click_bellreached the official success in 4 decisions and 12 model calls; the episode took about 4 minutes. GPU memory peaked at 8 GB for the simulator plus 8–10 GB for the perception services.
Choosing a task
--task <name>; --task-config demo_randomized (default, cluttered scenes
with random textures) or demo_clean; --seed picks a seed the qualification
accepted. The 50 tasks:
adjust_bottle beat_block_hammer blocks_ranking_rgb blocks_ranking_size
click_alarmclock click_bell dump_bin_bigbin grab_roller handover_block
handover_mic hanging_mug lift_pot move_can_pot move_pillbottle_pad
move_playingcard_away move_stapler_pad open_laptop open_microwave
pick_diverse_bottles pick_dual_bottles place_a2b_left place_a2b_right
place_bread_basket place_bread_skillet place_burger_fries place_can_basket
place_cans_plasticbox place_container_plate place_dual_shoes place_empty_cup
place_fan place_mouse_pad place_object_basket place_object_scale
place_object_stand place_phone_stand place_shoe press_stapler
put_bottles_dustbin put_object_cabinet rotate_qrcode scan_object
shake_bottle shake_bottle_horizontally stack_blocks_three stack_blocks_two
stack_bowls_three stack_bowls_two stamp_seal turn_switch
./scripts/ora run --env robotwin --task stack_blocks_two --task-config demo_clean --confirm-launch
Some tasks have few seeds the expert can solve. If the scan finds none, widen
it with --candidates 8.
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 RoboTwin installation: each pinned source, asset and seed-scan command, for a custom or multi-GPU host.