Read shared setup for host slots, model profiles, perception services and launch review. The default RoboTwin surface uses the ALOHA dual-arm robot, RGB-D perception and the shared planner; the expert and the seed scans use the vendored CuRobo planner. The simulator's check_success supplies the verdict.
At a glance
The dual-arm ALOHA-AgileX (two 6-DoF arms) in SAPIEN. 50 tasks, each in clean
(demo_clean) and randomized (demo_randomized) scenes; every task's
instruction and completion contract is on RoboTwin tasks.
Needs an NVIDIA GPU, CUDA for the planner, and per-seed qualification (below).
| Observation | head_camera and front_camera (fixed), left_camera and right_camera (on the grippers), RGB-D with per-observation calibration; both arms' TCP poses and measured jaw separation, the last action's receipt and the remaining budget |
| Action | Sparse per-arm motion proposals through the Supervisor (move_pose, move_to, rotations, set_gripper, recover_pose, wait_ticks), plus grasp_plan / grasp_execute and align_plan; poses name each arm's TCP |
| Speed | move_pose, move_to and the rotations take speed, speed_m_s and slow_approaching beside targets, as on every benchmark; every global route is timed at the TCP by the TOPP-RA + Ruckig worker under ORA_TIMING_PYTHON, which ora setup --env robotwin builds, and a hand turn more than 45 degrees past the request is refused before anything moves (motion speed) |
| Planning | move_pose's global routes, the reach check and align_plan plan on the shared planner under CUROBO_V2_PYTHON and CUROBO_V2_ROOT, which ora setup --env robotwin builds, in a collision world rebuilt per physical state from public depth with the other arm in it (planner configuration) |
| Success | RoboTwin's check_success, latched as eval_success, one verdict per round. check_task returns that boolean, never the predicate |
| Task prompt | The task's public instruction, plus its public completion contract |
Robot facts: ALOHA-AgileX · frames and units · policy I/O.
Source, interpreter and planner
export ORA_INSTALL_ROOT=/srv/ora
mkdir -p "$ORA_INSTALL_ROOT/sources" "$ORA_INSTALL_ROOT/envs"
git clone https://github.com/RoboTwin-Platform/RoboTwin.git "$ORA_INSTALL_ROOT/sources/RoboTwin"
git -C "$ORA_INSTALL_ROOT/sources/RoboTwin" checkout --detach c3ddfa8b97d5519efa828b075999bd0006778e5e
uv run --locked python -m ora.environment.backends.robotwin.vendored_install \
--robotwin-root "$ORA_INSTALL_ROOT/sources/RoboTwin"
git -C "$ORA_INSTALL_ROOT/sources/RoboTwin" rev-parse HEAD
# Expected: cf4872cc90538fd4bca761c8db912d8f65e49f2b
uv venv --python 3.10 --seed "$ORA_INSTALL_ROOT/envs/robotwin"
uv pip install --python "$ORA_INSTALL_ROOT/envs/robotwin/bin/python" \
torch==2.4.1 torchvision==0.19.1 --index-url https://download.pytorch.org/whl/cu121
uv pip install --python "$ORA_INSTALL_ROOT/envs/robotwin/bin/python" \
-r "$ORA_INSTALL_ROOT/sources/RoboTwin/script/requirements.txt" \
numpy==1.26.4 torchvision==0.19.1 setuptools==69.5.1
git clone https://github.com/NVlabs/curobo.git "$ORA_INSTALL_ROOT/sources/curobo"
git -C "$ORA_INSTALL_ROOT/sources/curobo" checkout --detach d64c4b005459db10c5dd867d8b30a87d5bda9bdb
# CUDA_HOME must select a toolkit compatible with this Torch environment.
uv pip install --python "$ORA_INSTALL_ROOT/envs/robotwin/bin/python" \
-e "$ORA_INSTALL_ROOT/sources/curobo" --no-build-isolation
uv pip install --python "$ORA_INSTALL_ROOT/envs/robotwin/bin/python" warp-lang==1.12.0
export ROBOTWIN_ROOT="$ORA_INSTALL_ROOT/sources/RoboTwin"
export ROBOTWIN_PYTHON="$ORA_INSTALL_ROOT/envs/robotwin/bin/python"
export CUROBO_ROOT="$ORA_INSTALL_ROOT/sources/curobo"
export CUROBO_PYTHON="$ROBOTWIN_PYTHON"
The installer creates a deterministic source commit. Do not edit the planner to bypass the revision gate. The original identity text retained inside the installer is part of that commit's reproducibility.
Assets
Download background_texture.zip, embodiments.zip and objects.zip from TianxingChen/RoboTwin2.0, selecting and recording an immutable dataset revision. Training checkpoints and demonstration trajectories are unnecessary. Record archive SHA-256 values, extract all three into $ROBOTWIN_ROOT/assets, then regenerate absolute robot paths:
(cd "$ROBOTWIN_ROOT" && "$ROBOTWIN_PYTHON" script/update_embodiment_config_path.py)
uv run --locked ora doctor --env robotwin
The assets/background_texture/unseen/ pool must be complete. Moving the installation requires regenerating embodiment paths. SAPIEN requires working Vulkan libraries, not merely a visible CUDA GPU.
Qualify one seed
On your registered robotwin-gpu0 slot, prepare a bounded scan for one task. This runs the upstream expert, not the language model:
./scripts/ora qualify robotwin prepare \
--robotwin-root "$ROBOTWIN_ROOT" --robotwin-python "$ROBOTWIN_PYTHON" \
--tasks adjust_bottle --slots robotwin-gpu0 \
--accepted-per-task 1 --candidate-count 3 --task-config demo_clean
./scripts/ora qualify robotwin qualify \
--plan PATH_TO_PRINTED_PLAN --confirm QUALIFY-ROBOTWIN:PRINTED_HASH
The immutable plan gives the estimate and confirmation hash. Candidate probes can take minutes (the three-candidate adjust_bottle scan above finished in about 7 minutes on one data-center GPU lane, under its 28.5-minute estimate); this bounded scan is separate from a full 50-task qualification. Preparation and qualification records are written under runtime-v1/robotwin-evaluation/ in the outputs root (ORA_OUTPUTS_ROOT when set, otherwise <workspace>/outputs); this namespace is fixed. qualify prints the path of the resulting seed-manifests/adjust_bottle.json, which names accepted seeds and binds the source/configuration. Select a seed actually present in it. Reuse still-valid evidence instead of rescanning on every launch. See commands.
Compose and launch
After qualifying perception services, substitute their report path, the seed manifest and an accepted seed:
./scripts/ora spec robotwin compose adjust_bottle \
--seed ACCEPTED_SEED --task-config demo_clean \
--qualification PATH_TO_SEED_MANIFEST \
--experiment robotwin-example --resource-profile robotwin-gpu0 \
--service-report PATH_TO_READINESS_REPORT --cohort robotwin-example-g001
Use the printed path in the review/launch commands. The result contains the official boolean verdict; completion alone does not imply success. demo_randomized needs qualification for that configuration. The composer inherits the current default tools/prompts and an 80-decision, 200-model-call, 10,000-sim-step, six-hour ceiling. Actual duration and tokens are reported in the episode evidence.
Free play
ora play --env robotwin plays an audited seed of an official task without its
success check, or, with --scene-set tabletop-v1 or --scene-set household-v1
(or --scene naming one of their scenes), one of ORA's own tables with many
objects: three rebuilt prototypes, or a breakfast table and a kitchen counter
whose objects have household masses. Audit a scene set on your RoboTwin slot
before the first session, and again after a change to its scene modules:
./scripts/ora play audit --env robotwin --scene-set household-v1 --seeds 5 --resource-profile robotwin-gpu0
./scripts/ora play --env robotwin --scene kitchen --provider PROFILE --service-report PATH_TO_READINESS_REPORT --confirm-launch
A scene of a set has its own operator camera for you (1280 × 720, from your
side of the table at three quarters, never one of the agent's cameras), which
/view and the web page show. Its execution capture records that camera beside
the head and front cameras at 30 fps of simulated time, and after each action
the scene keeps running, the arms holding their last command, until its objects
rest (at most 2 s), so a knocked-over object is seen falling. The wall time of
a motion grows about twofold; see free play.
Official tasks in free play keep RoboTwin's observer camera and the 10 fps capture.