A preprint in computer vision reports that UniAfford, a single model for 2D and 3D affordance grounding, achieved the strongest listed results in several cross-dataset tests. It led the listed non-reasoning methods on gIoU and cIoU in a 2D transfer benchmark, and the authors say it beat every evaluated zero-shot baseline on four measures in a 3D transfer test.
A common language for two data types
UniAfford is built around a multimodal large language model used as a shared semantic hub. It routes contextual response states to text, 2D-affordance and 3D-affordance branches, and those routed states condition SAM-style and SONATA-based decoders. The stated aim is to learn transferable 2D and 3D semantics from heterogeneous supervision.
Training combines language modeling, 2D and 3D affordance prediction, and supervision for the token-routing process. Dense objectives then supervise selected routed representations through the corresponding decoders.
The UniAfford-Data comparison lists 162 objects, 162 affordances, 25,000 2D samples and 69,000 3D samples. These figures are reported by modality, rather than as one single non-overlapping total.
The main test used a mixed-training, out-of-distribution zero-shot protocol, with no target-specific training or fine-tuning. A second modality-isolated protocol separately trains and evaluates the 2D and 3D branches.
Across the benchmarks
On AGD20K, the out-of-distribution 2D test, UniAfford reported 27.52 gIoU and 25.22 cIoU, along with 6.73 on P50 to P95, 19.88 on P50, 4.30 KLD and 0.37 SIM. The paper says those results led the listed non-reasoning methods on gIoU and cIoU and gave UniAfford the highest listed SIM.
On GEAL*, the 3D transfer test, it reported 83.55 AUC, 14.67 mIoU, 0.565 SIM and 0.102 MAE. The authors say this exceeded every evaluated zero-shot baseline on the reported measures and surpassed the reference models on AUC and SIM.
The reported numbers are point estimates. No uncertainty estimates are reported alongside the comparisons, so the paper does not show how variable the measured differences might be.
In the modality-isolated ReasonAff evaluation, UniAfford led all reported 2D metrics: 71.19 gIoU, 73.63 cIoU, 80.94 P50 and 55.89 P50 to P95.
In the 3D-only PIAD comparison, UniAfford recorded 14.25 mIoU versus DAG's 9.73, a reported 4.52-point difference, with the best AUC and MAE and tied-best SIM. On PIADv2, it outperformed GREAT across all four reported metrics.
What the design tests show
A controlled ablation compared the full design with fixed-anchor routing and single-branch training. The full model scored 68.79 2D gIoU, versus 62.28 for fixed-anchor routing and 41.41 for 2D-only training. Its 3D mIoU was 34.56, compared with 17.46 for fixed-anchor routing and 30.07 for 3D-only training. These are descriptive score differences, not randomized causal estimates of what any one component contributed.
Another ablation replaced similarity-based decoder coupling with prompt-style alternatives. The 2D variant scored 36.97 gIoU and the 3D variant scored 14.31 mIoU; the authors describe both as reductions relative to the similarity-based full design.
A language-head diagnostic also recorded broader decoded-token category counts for the jointly trained model: 39 categories for ReasonAff-only training, compared with 238 on ReasonAff, 398 on GEAL and 426 on AGD20K for UniAfford-Data. The readouts included object, part and interaction-related terms. The paper notes that these counts are not sample-size-controlled measures of semantic richness.
Where the evidence stops
The benchmark advantage has a narrower meaning than robot performance. The study does not evaluate closed-loop robotic manipulation. It also identifies a lack of instance-level spatial correspondence in its semantic pairing: object-affordance labels connect different physical instances but do not establish their exact spatial relationship. Large pretrained backbones add training and inference costs.
The evidence therefore supports comparative claims under the stated offline benchmark protocols, not a demonstration that a robot can manipulate objects more successfully in the real world.
UniAfford is labeled a preprint, identified as arXiv:2609.37264v1 and dated 29 September 2026.
Paper data and sources
Original title: UniAfford: Token-Routed Multitask Learning for Generalizable 2D-3D Affordance Perception
Authors: Yuhao Liu, Yiming Zhong, Hanqing Wang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-09-29
DOI: Not available
Original paper · Full text