Task-Conditioned Scene-level Affordance Grounding for Real-Time Manipulation
A2A-AffordGen — our agent-assisted annotation pipeline — is now available in
A2A-AffordGen/, and the A2A-Bench dataset card is up (data hosting in progress). A2A-GroundingModel and A2A-Policy are still being cleaned up for release. ⭐ Star/Watch the repo to get notified.
Affordance2Action (A2A) is a benchmark-centered learning framework for scene-level, task-conditioned part affordance grounding. Task-conditioned manipulation requires grounding instructions to task-relevant functional parts rather than object categories — a setting that is scene-dependent and often one-to-many: the same object may afford different interactions across tasks, while a single task may correspond to multiple valid functional regions.
At its core is A2A-Bench, a manipulation-oriented benchmark that covers both single-region and multi-region instruction correspondences in everyday scenes. To build it at scale, we develop A2A-AffordGen, an agent-assisted annotation pipeline. A2A-Bench's supervision enables a real-time grounding model (A2A-GroundingModel) and a manipulation policy (A2A-Policy).
- 🌐 Project page: https://arc-l.github.io/a2a/
| Component | Description | Status |
|---|---|---|
| A2A-AffordGen | Agent-assisted annotation pipeline | ✅ Released |
| A2A-Bench | Scene-level, task-conditioned, one-to-many affordance benchmark | 📄 Dataset card · data hosting in progress |
| A2A-GroundingModel | Real-time task-conditioned part grounding (SAM3-based) | 🚧 Coming soon |
| A2A-Policy | Manipulation policy guided by affordance priors | 🚧 Coming soon |
@article{liu2026a2a,
title = {Affordance2Action: Task-Conditioned Scene-level Affordance Grounding for Real-Time Manipulation},
author = {Liu, Litao and Han, Yifan and Yi, Pengfei and Yu, Wenbo and Wang, Hanqing and
Du, Haoran and Yuan, Enze and Yuan, Zilin and Feng, Ruiding and Liu, Michael and
Zhang, Qi and Yu, Jingjin},
year = {2026}
}