An arXiv preprint reports GOAG, a robotic grasp planner designed to generate grasp contacts without relying on object-specific grasp databases. It takes gripper configurations and contact points as its starting point and learns object-agnostic contact distributions, with the stated aim of handling unseen object shapes. The authors present this as a shift from object-centric or grasp-centric planning toward a gripper-centric view.
In the paper’s main MultiDex simulation comparison, GOAG reported an average success rate of 86.93% across Barrett, Allegro and ShadowHand. These are results from the paper’s stated evaluations; they do not establish universal performance across arbitrary objects or grippers.
The document is labeled arXiv:2608.19759v1 and dated 20 Aug 2026. It is a preprint, so these findings represent the authors’ reported evaluation rather than a settled performance standard.
A planner built around the hand
GOAG’s pipeline combines a conditional variational autoencoder, or CVAE—a model used here to generate contact patterns—with BPS workspace encoding and a PointNet++ link mapper. In the paper’s description, these components learn object-agnostic contact distributions from gripper configurations and contact points, then connect candidate contacts to the gripper’s links. The approach also includes a force-closure estimate, a geometric screen for contact feasibility.
The training data were synthetic. The authors uniformly sampled 10,000 kinematically valid gripper configurations, selected six grasp types and sampled 50 contact-point sets for each configuration. The process produced 3,000,000 labeled point clouds.
The paper reports that creating this synthetic dataset took approximately one GPU hour on an Nvidia RTX 4090. It compares that with 1,400 GPU hours reported for prior work using an Nvidia A100. No uncertainty or repeated timing statistics were reported for this comparison.
How the test defined success
Grasp success was assessed in Isaac Gym under a defined disturbance sequence. Sequential external forces were applied along the positive and negative x, y and z directions for one second each, and a grasp counted as successful when the object’s deviation stayed below 2 centimetres. The reported success percentages therefore refer to that specified simulated force test.
The MultiDex comparison covered three grippers, 100 inference runs per method, a batch size of 10 and 10 test objects. GOAG was evaluated against several baseline planners, with DFC included in the table’s efficiency comparison.
GOAG reported 87.40% success for Barrett, 93.20% for Allegro and 77.90% for ShadowHand, producing the 86.93% average. Allegro had the highest reported rate and ShadowHand the lowest, so the headline average combines results from multiple grippers rather than describing a single hardware setup.
Fast generation and a component check
The table reports efficiency entries of 0.18, 0.19 and 0.20 for GOAG, compared with DFC entries greater than 1800. It also lists GOAG diversity entries of 0.0479, 1.401 and 0.3170. The supplied analysis describes GOAG as faster than DFC in this comparison, but gives no uncertainty estimates for the entries.
An ablation removed force-closure estimation from the pipeline. The resulting row reported 86.30%, 91.20% and 74.70% for the three grippers, with an 84.07% average, compared with 86.93% for full GOAG. This is a comparison between reported rows; the figures alone do not establish that the removed component caused the difference.
A broader comparison, with a caveat
GOAG was also tested across five benchmarks containing 3,438 objects. The reported benchmark success rates were 43.07%, 49.51%, 77.90%, 37.37% and 62.13%, with a 53.97% average. The authors described that average as the second-highest in their comparison.
That ranking comes with an important qualification. The paper says GOAG was trained once on Shadow Hand, while competing methods were retrained for each dataset. The comparison therefore uses different training regimes, so it should be read as a cross-dataset result under the paper’s setup rather than a fully matched head-to-head test.
The physical demonstration was limited
The authors also report a real-robot demonstration: an Allegro Left Hand mounted on a 7-DoF arm successfully grasped 11 YCB objects. That shows the pipeline was connected to physical hardware in the reported experiment, but it does not establish broad reliability across hardware, objects or repeated trials.
A key technical caveat concerns force closure, the geometric screen used to assess whether a set of contacts is feasible. The estimate does not account for gravity or object weight. The analysis also notes that contacts judged valid by this screen can later be unreachable or obstructed, making final physical validation necessary.
What the result still leaves open
The reported percentages and table comparisons are descriptive. The supplied analysis reports no inferential tests, confidence intervals or formal uncertainty estimates, and the real-robot demonstration has no repeated-trial statistics. The paper therefore reports outcomes without quantifying how much the figures might vary.
The remaining tests are practical ones: how the method performs on larger objects, whether the cross-dataset ranking persists under matched training and preprocessing, and whether simulated success carries over to broader hardware and repeated real-world trials. The authors nevertheless interpret GOAG as a generalizable and competitive approach that avoids object-specific grasp databases and transfers to a real robot.
The authors state that code and videos are available on the project website, providing material for others to inspect the implementation and demonstrations.
Paper data and sources
Original title: GOAG: Generative and Object-Agnostic Grasp Planner for Dexterous Robotic Manipulation
Authors: Julien Merand, Boris Meden, Mathieu Grossard, Liming Chen
Journal/Repository: 2026 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text