GraspXL
ETH / GraspXL research team
Generates grasping motions with specified approach directions and object objectives, with examples for several hand models.
What you can explore
Study how grasp objectives and object geometry affect generated motion; inspect the linked generated-data release.
Before you use it
RaiSim requires its own activation and license. Generated grasps are simulation results. Different embodiments and mesh collections have distinct coverage and asset conditions.
Original contributors
Hui Zhang, Sammy Christen, Zicong Fan, Otmar Hilliges and Jie Song
License / access: Work and dataset: CC BY-NC 4.0 · RaiSim and object assets have separate terms
GraspXL: inspect one grasp before training at scale
- Environment
- Separate visualization environment; RaiSim for policy training
- Embodiment
- Choose the exact released hand and motion collection
What you need
Decide whether you want to examine data or reproduce the policy. The standalone viewer is a smaller first task than setting up training. Its environment file pins Linux/CUDA dependencies; do not assume it installs unchanged on a Mac. The training repository uses a different environment and requires RaiSim access. Keep the viewer, training code and downloaded assets in separate, versioned environments.
First useful check
Choose one object and one embodiment. Explain where the hand starts, when contact occurs and how the object moves. Save a short inspection report before investing in training.
A small first experiment
Choose the viewer route and pin its environment
Start with the author’s visualization repository. Follow its environment file and asset instructions, and record both the viewer and data revisions. The repository has separate scripts for different hands and tabletop collections. Select a script that actually matches your downloaded sequence, rather than renaming a file from another embodiment.
Author’s standalone viewer →Match the motion, mesh and hand model
For the original Allegro viewer, check that the sequence and object mesh are in the paths expected by its script, and that its Allegro URDF is present. The script reads hand translation, rotation and pose separately; its pose vector includes root coordinates as well as finger joints. Do not send that vector to a motor driver. MANO examples additionally need separately licensed MANO assets. Save this standard-library check as check_graspxl.py and run python check_graspxl.py /path/to/GraspXL_visualization allegro_WineGlass wineglass. It checks the original Allegro layout without loading serialized data or importing the viewer. Use exact names (case-sensitive on Linux) and omit file extensions. This is not the tabletop layout; a PASS verifies presence only.
Inspect the actual Allegro loader →from pathlib import Path import sys if len(sys.argv) != 4: raise SystemExit("Usage: python check_graspxl.py VIEWER_ROOT SEQUENCE OBJECT") root = Path(sys.argv[1]).expanduser().resolve() sequence, obj = sys.argv[2:] for name in (sequence, obj): if not name or name in (".", "..") or any(c in name for c in "/\\"): raise SystemExit("Use a basename without directories or extensions") paths = { "motion": root / "data/GraspXL/recorded" / (sequence + ".npy"), "object": root / "data/GraspXL/object_mesh" / (obj + ".obj"), "hand": root / "data/body_models/allegro/allegro.urdf", } missing = [] for label, path in paths.items(): exists = path.is_file() print(("FOUND" if exists else "MISSING"), label, path) if not exists: missing.append(label) if missing: raise SystemExit("Missing inputs: " + ", ".join(missing)) print("PASS: filenames exist; contents, URDF mesh references and motion are NOT checked")Inspect a single released sequence
After installing the viewer and obtaining its required assets, inspect the options below from the viewer repository root. Follow the author’s example with your matching sequence and object. Pause at approach, first contact and lift; record intersections, sudden pose jumps and object-scale mismatches. Only load NumPy/pickle objects from a trusted source.
Original data and collection descriptions →python scripts/visualizer_allegro.py --helpDesign a separate policy evaluation
When the data inspection works, move to the training repository’s matching hand environment and released checkpoint. Keep training and evaluation objects separate, record reset conditions, and define failure before running trials. Compare the same task and embodiment across methods; a saved successful motion is not an unbiased success-rate sample.
Build a reproducible grasp evaluation →
What to record
- Record viewer/training commits, collection name, sequence ID, object ID, mesh scale and hand-model revision.
- Save an inspection image and note whether the pose includes root coordinates, finger joints or MANO parameters.
- For a later policy test, report the checkpoint, trial count, resets, success rule and failures separately from playback.
Common problems and what to check
- File exists but the viewer cannot find it
- Run the viewer from its repository root. The reviewed original loader builds relative paths and adds .npy and .obj itself; keep the sequence and object basenames separate. Tabletop collections use nested paths and their own viewer script.
- Mesh and grasp do not line up
- Confirm that motion and geometry come from the same collection and use its scale. Substituting an original textured mesh without matching geometry can invalidate the comparison.
- The viewer environment will not resolve
- The published environment contains platform-specific pins. Check its operating-system and GPU assumptions first; installing the training environment does not satisfy the viewer’s dependencies.
- A different hand has the same vector length
- Inspect named joints, root pose and rotation conventions. Matching array sizes do not establish matching actions.
Where it stops
Playing a saved trajectory does not test a policy in closed loop. Visual contact is not proof of a stable physical grasp. Data, simulator, hand assets and code have separate terms; check the linked licenses before commercial use. This guide does not establish compatibility with a physical hand.
Keep exploring
DexGraspNet
Synthetic dexterous grasp configurations generated for a Shadow Hand and validated in simulation.
RobustDexGrasp
Teacher and student policies for grasping unfamiliar objects using an Allegro hand and UR5 platform.
Editorial reference · checked 2026-09-22