Venue: ICML 2026 (Poster) Category: Benchmark Affiliations: Authors: Xingyu Zhu, Wenshuo Han, Zhouyu Wang, Yuran Wang, Ruihai Wu, Hao Dong, Fan Tang, Hechang Chen, Hyung Jin Chang, Yixing Gao

Robotic manipulation of flat objects (paper, cloth, cards, thin plates) is hard because such objects often start in ungraspable configurations and exhibit "strong variations in object geometry and material." Existing methods rely on heuristic pre-manipulation and are "often evaluated in closed settings with limited generalization," leaving no standardized way to measure progress.

FlatLab couples a manipulation framework with a benchmark. The framework decouples manipulation into a strategy generator and an action execution module:

• Strategy generator. Predicts appropriate manipulation strategies directly from object point clouds by learning "strategy-centric, object-invariant representations via simulated data transformation and contrastive learning." This makes strategy prediction generalize across object instances rather than memorizing geometry.
• Action execution module. Conditioned on the predicted strategy, it "decomposes long-horizon manipulation into reusable action primitives and dynamically composes them to generate stable trajectories."

The accompanying FlatLab benchmark provides high-fidelity physical simulation of diverse rigid and deformable flat objects, automated multi-modal data collection, and standardized task definitions and evaluation protocols.

flowchart LR
    PC[Object point cloud] --> SG[Strategy generator<br/>object-invariant representation<br/>contrastive learning]
    SG --> STR[Predicted strategy]
    STR --> EX[Action execution module]
    EX --> PRIM[Reusable action primitives]
    PRIM --> TRAJ[Dynamically composed<br/>stable trajectory]
    BENCH[(FlatLab benchmark<br/>rigid + deformable flat objects<br/>multi-modal data, std protocols)] -.evaluates.-> TRAJ

The paper does not report specific numbers in the available abstract. It states that "Experiments conducted in FlatLab demonstrate that our approach generalizes effectively to unseen objects and categories, outperforming existing baselines."

Flat-object manipulation is an underserved but practically important slice of robotic manipulation (document handling, garment and sheet manipulation, assembly of thin parts). By pairing a generalizable strategy/execution framework with a high-fidelity, standardized simulation benchmark covering both rigid and deformable items, FlatLab gives the 2026 community a common yardstick for a problem that previously lacked one.

• ICML 2026: https://icml.cc/virtual/2026/poster/66662

← Back to ICML-2026