CoToGrasp: Contact-Topology-Conditioned Dexterous Grasp Synthesis via Canonical Workspace Learning
Julien Merand, Boris Meden, Liming Chen, Mathieu Grossard
ECCV)
cs.RO, cs.AI
2026-08-20
CoToGrasp generates 21-topology grasps with no object meshes in training. DexGraspNet precision success is 30.3% vs Dexonomy 10.5%; topology compliance 17.18% vs 14.28%.
Dexterous grasp planners have been scored on one question: can this object be held. Force-closure optima collapse into enveloping power grasps. Downstream work often needs a different contact pattern: a fingertip pinch, a tool handle, a finger gait. Unconditioned generators waste samples hoping to draw those modes.
Conditioning on human grasp taxonomies usually means labelling grasp types onto specific object meshes, which is expensive enough to stall the pipeline. The same group's earlier GOAG paper showed that contact can be learned in the gripper frame alone. CoToGrasp attaches a functional condition to that idea. The question becomes how to grasp under a named contact topology, not whether a grasp exists.
Contact topologies come from Gonzalez's hand-centric taxonomy: 21 templates and 22 anatomical zones, defined on which hand surfaces must touch, not on object shape. Training is object-agnostic. Ten thousand kinematically valid joint configurations, each paired with all 21 templates, yield 210,000 samples, no object meshes and no physics.
The geometric bridge is a canonical feature workspace in the gripper frame. A DGCNN extracts local point features; a kNN gated by aligned distance aggregates them onto a fixed set of workspace basis points. Training sees gripper clouds; inference sees the object transformed into the gripper frame with inverted normals. Both land on the same spatial tokens. A Transformer models co-occurrence of contacts, a Set Transformer compresses a global descriptor, and a CVAE samples local intra-topology variation. Macro-level multimodality is already resolved by the topology label, so a unimodal Gaussian prior is enough.
At test time a cascade filters candidates. Predicted and template masks may differ by at most one contact zone. Survivors go through an approximate force-closure check; failures resample the latent code, up to 20 times. An energy step then seats the gripper's active surface on the predicted zones, with a 5 mm repulsive field on unused fingers and palm so idle phalanges do not smear a precision grasp into a power grasp. The global 6D pose is an external prior, not a network output.
Against unconditioned generators CoToGrasp looks weaker on physical success, because it is forced to emit precision pinches that geometry does not favour. Semantic entropy is the highest of the group, and so is speed.
| Method | Success rate | Semantic entropy HTC | sec/grasp |
| DFC | 72.15 | 0.7389 | >1800 |
| GenDexGrasp | 71.15 | 0.5956 | 14.65 |
| DRO-Grasp | 63.30 | 0.6504 | 1.72 |
| GOAG | 77.90 | 0.6527 | 0.20 |
| CoToGrasp | 36.94 | 0.83 | 0.11 |
Top-5 average success is 56.18%. Against the taxonomy-aware baseline Dexonomy on DexGraspNet, mean success is 26.72% vs 21.13% and topology compliance 17.18% vs 14.28%. Precision-class success is 22.71% vs 12.36%; the M2 pinch is 30.3% vs 10.5%. Dexonomy's 60.5% and 37.2% on M11 and M21 are inflated by collapse into unchecked enveloping grasps. Dropping the label-consistency filter sends compliance back to 14.45%, Dexonomy's band.
On a UR10 with an Allegro right hand, YCB objects can be closed and held under named topologies. M6 is suppressed because the Allegro has no little finger. On non-convex objects CoToGrasp retains 58.72% of its success against Dexonomy's 37.42%; on severe concavities (about 4% of the set) compliance is 19.17% vs 8.94%.
The data bottleneck for functional grasping is sidestepped: taxonomy conditions no longer have to be bound to object meshes. For humanoid stacks that take high-level intent from an LLM, pinch, handle and envelope become controllable channels rather than rejected samples from a power-grasp prior. At 0.11 s per grasp, resampling online is realistic.
This is a directional, incremental step. Physical success sits below unconditioned planners because the protocol forces hard topologies. In the fair Dexonomy comparison the win is precision classes and anti-collapse, while absolute compliance remains in the teens. "Grasp as instructed" is still some way off under this strict metric.
The headline 36.94% success is not a deployment number; Top-5 only reaches 56%. Compliance drops after physics: fingers slide a few millimetres for force closure, and the asymmetric Tversky score reclassifies a valid pinch as a neighbour type. 6D approach poses are sampled by a heuristic outside the network. Real precision grasps expose position-control mismatch: fingers close at uniform speed, leading contacts knock the object off the planned wrench. Tabletop scenes inflate grasps that would need a second hand or a constrained presentation. The Allegro is four-fingered, so five-finger templates are cut. Transfer of the taxonomy onto non-anthropomorphic hands has no numbers.