GOAG: Generative and Object-Agnostic Grasp Planner for Dexterous Robotic Manipulation
Julien Merand, Boris Meden, Mathieu Grossard, Liming Chen
IROS)
cs.RO, cs.AI
2026-08-20
GOAG trains a CVAE on gripper kinematics and injects objects only at inference. MultiDex mean success is 86.93% across three hands; dataset generation takes one GPU-hour.
Learned multifingered grasp planners are trained on finite object–gripper pairings. A shape outside that set takes the contact map and the joint solution down with it. Analytical planners generalise and search too slowly. Data-driven ones are fast and write object geometry into the weights.
CEA List flips the contact definition. At the contact patch the gripper and the object share the same surface geometry, so the contact set should look the same from either side. The hand's feasible contact distribution can be learned first; object geometry enters only at inference. GOAG is a generative planner that is object-agnostic during training.
All training data comes from the gripper. Ten thousand valid joint configurations, six everyday grasp types adapted from the literature (covering 92.5% of machinist and 96.2% of housemaid working time), fifty contact samples per configuration inside the admissible regions: three million labelled clouds. Dataset generation takes about one GPU-hour on an RTX 4090. Object–grasp libraries in the GenDexGrasp line reported 1,400 A100 GPU-hours.
Geometry is encoded with a Basis Point Set of 8,192 points laid over the gripper workspace. Each basis stores an aligned distance into the input cloud. A CVAE with a 128-D latent learns a continuous contact-likelihood field on those bases; a parallel PointNet++ maps contact points onto phalanx links. At inference the novel object is transformed into the gripper frame, encoded with the same BPS, and decoded from a prior sample z. Bases above τ=0.8 become contact targets.
Contacts clustered by phalanx go through a force-closure pre-check (Coulomb μ=0.3, no gravity). Failures resample, at most 20 times. Survivors are refined by an energy on contact fit, object penetration, self-collision and joint limits. Global pose is not a network output: palms are sampled on the convex hull dilated by 110%, facing against the hull normal.
On MultiDex (ten objects, three grippers), Isaac Gym pushes ±xyz for one second each; displacement under 2 cm counts as success. GOAG averages 86.93%, ahead on every hand.
| Method | Barrett | Allegro | Shadow | Mean | sec/grasp (Shadow) |
| DFC | 83.10 | 82.71 | 72.15 | 79.32 | >1800 |
| GenDexGrasp | 70.26 | 71.48 | 71.15 | 70.96 | 14.65 |
| DRO-Grasp | 78.30 | 75.80 | 63.30 | 72.47 | 1.72 |
| GOAG (no FC) | 86.30 | 91.20 | 74.70 | 84.07 | 0.15 |
| GOAG | 87.40 | 93.20 | 77.90 | 86.93 | 0.20 |
Without the force-closure filter the mean is still 84.07%, so the contact prior already carries usable mechanics. DRO-Grasp can look faster on a single grasp and then scales linearly; GOAG vectorises candidates into one energy, which holds up in bulk.
Cross-dataset generalisation uses the Shadow Hand on 3,438 objects. Baselines retrain per dataset; GOAG trains once on hand kinematics: DexGraspNet 43.07, UniDexGrasp 49.51, MultiDex 77.90, RealDex 37.37, DexGRAB 62.13, mean 53.97%. Second of five, 4.5 points behind DGA at 58.48%, which was retrained on each set. A 7-DoF arm with an Allegro left hand lifted 11 YCB objects. There is no per-object real-world success table.
A paired object–grasp database is no longer the ticket into learned dexterous grasping. New objects do not require thousands of GPU-hours of labelled simulation; a new gripper does, because the contact manifold is hand-specific. For a line that keeps adding SKUs, that asymmetry is the product: zero-shot on objects, one-shot modelling on the hand.
The numbers are a solid increment. The MultiDex lead is clean. The five-set mean still sits under a method trained on each benchmark. Object-agnostic is not gripper-agnostic, and functional intent is not in the generator. The same authors' later CoToGrasp hangs contact-topology conditions on this skeleton.
There is no standalone limitations section. Pose sampling assumes the object fits inside the gripper workspace; large objects degrade, and the suggested fix has no numbers. Force closure uses contact barycenters and ignores gravity, so it is not the Isaac success metric. The train–test domain shift (gripper geometry in, object geometry in) is bridged by BPS with limited ablation. Eleven real objects are a demo, not a controlled study. Six grasp types cover frequent daily actions; corner topologies are outside the model.