GOAG: Generative and Object-Agnostic Grasp Planner for Dexterous Robotic Manipulation

The central bet in GOAG is geometric: at the moment a multi-fingered hand closes on an object, the contact surface looks identical whether you describe it from the object's frame or the gripper's. That equivalence is not just a curiosity — it is a training strategy. By flipping to the gripper's canonical frame, the authors build a planner that never sees an object during training at all.

The mechanics are straightforward once the perspective shift is accepted. A Conditional Variational Autoencoder learns the distribution of feasible contact zones purely from the gripper's own kinematics and a six-type grasp taxonomy adapted from human-hand research. Training data is generated procedurally: sample a joint configuration, pick a grasp type, scatter contact points within the admissible regions. The whole dataset for one gripper took roughly one GPU hour on an RTX 4090. A comparable object-specific dataset from prior work reportedly required fourteen hundred GPU hours on A100s — a difference that reflects not just compute but the fundamental cost of needing real object geometry at all.

At inference, the object point cloud is transformed into the gripper's workspace frame and encoded with a Basis Point Set. The trained decoder then predicts which parts of that workspace are likely contact zones, a PointNet++ assigns each predicted point to a specific finger link, and a force-closure check filters unstable configurations before a final joint optimization refines the pose. The resampling loop caps at twenty iterations to avoid spinning on poorly conditioned object poses near workspace boundaries — a practical acknowledgment that the approach is not universally robust.

On the Multidex benchmark against Barrett, Allegro, and Shadow Hand grippers, GOAG outperforms baselines including GenDexGrasp and DRO-Grasp on average success rate, despite those methods being trained on the same dataset's objects. The efficiency story is nuanced: DRO-Grasp is faster per single grasp, but its time scales linearly with batch size because each grasp is optimized independently. GOAG vectorizes the energy minimization across all candidates simultaneously, so its advantage grows with batch size. Across five heterogeneous benchmarks using the Shadow Hand, GOAG achieves the second-highest average success rate — behind methods that were retrained on each dataset individually, while GOAG was trained exactly once.

The real-robot results are modest in scope: eleven YCB objects grasped with an Allegro hand on a seven-degree-of-freedom arm. That is enough to confirm sim-to-real transfer is not broken, not enough to characterize failure modes at scale.

The genuine limitation the paper names is pose sampling: the convex-hull strategy works well for objects whose volume fits inside the gripper's workspace, but larger objects need a different sampling approach that GOAG does not yet provide. The force-closure step also operates on simplified contact barycenters, not the full geometry, so configurations that pass that check can still fail during joint optimization when the actual object shape intervenes.

What GOAG demonstrates is that the object-specificity baked into most grasp datasets is not a requirement — it is a historical artifact of framing the problem from the object's side.