AAAI Poster · ReAct

ReAct: Reward-informed Autoregressive Decision CAD Transformer

ReACT recovers editable CAD modeling sequences directly from point clouds. It frames CAD reconstruction as an offline, reward conditioned decision process, using scaffold aware geometric states and dense shape rewards to predict accurate and concise sketch extrude commands.

Paper (PDF) Demo / Video About / Home

Authors

Yijie Ding1, Yang Liu2, Haobo Jiang1, Jianmin Zheng1†
Affiliations:
1Nanyang Technological University
2Beijing Institute of Technology
†Corresponding author

Highlights

  • Formulates CAD sequence reconstruction from point clouds as an offline, reward conditioned decision process.
  • Introduces a scaffold aware geometric state that fuses point cloud features with an incrementally built surface scaffold.
  • Designs a Local Barrel Point (LBP) guided dense reward that balances surface fidelity and command efficiency.
CAD Sequence Modeling Reinforcement Learning Reverse Engineering

Quick Links

  • ArXiv: tbd
  • Slides: tbd
  • Poster: tbd
  • Contact: yijie002 at ntu.edu.sg

Abstract

Reconstructing CAD modeling sequences from point clouds is difficult, especially for shapes with complex geometry and topology. ReACT addresses this by casting CAD reconstruction as an offline, reward conditioned decision process in which a transformer predicts sketch extrude commands from scaffold aware geometric states and dense, shape informed rewards. The model uses Local Barrel Point guided rewards and point command fusion to align surface fidelity with command efficiency, and achieves lower Chamfer Distance and invalid rates than imitation based baselines on DeepCAD and Fusion360.

Problem Recovering procedural CAD sequences from raw point clouds is challenging because point sets lack semantic structure and existing methods are trained only by imitation, which limits their ability to plan over long design histories and complex topology.
Method ReACT models each sketch extrude operation as an action in a Markov decision process, builds a scaffold aware geometric state from point cloud features and accumulated LBPs, and uses a return to go conditioned transformer to autoregressively predict the next CAD command.
Impact On DeepCAD and Fusion360, ReACT attains lower Chamfer Distance and invalid rates than DeepCAD, TransCAD and a Text2CAD style autoregressive baseline, while often producing shorter yet structurally faithful command sequences.

Figures & Visuals

Replace the placeholders below with your architecture diagram, qualitative trajectories, and quantitative comparisons.

Figure 1 Architecture diagram placeholder Drop in a 16:9 schematic or animated GIF.
Figure 2 Trajectory showcase Show reasoning steps, actions, and tool outputs.
Figure 3 Benchmark results Bar/line charts comparing baselines vs. ReAct.

Results Snapshot

Summarize the strongest numbers and key takeaways. You can replace the text with a small table or bullets.

State-of-the-art ReAct surpasses baseline agents by +X% on benchmark A and +Y% on benchmark B.
Efficiency Achieves Z% fewer tool calls with similar or better accuracy, reducing runtime and cost.
Robustness Maintains stable performance under noisy inputs and adversarial distractors.

Citation

Replace the fields below with your final citation once the AAAI version is confirmed.

@inproceedings{ding2026react,
        title = {ReACT: Reward informed Autoregressive Decision CAD Transformer},
        author = {Ding, Yijie and Liu, Yang and Jiang, Haobo and Zheng, Jianmin},
        booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence},
        year = {2026},
        note = {To appear},
        url = {https://openreview.net/forum?id=6232}
}