BrickFlow uses SE(3) flow models to reconstruct buildable LEGO assemblies from images
Peter Kulits · hf · 2026-11-17
Peter Kulits (Inria/ENS) and Cordelia Schmidt (Max Planck Institute) present BrickFlow, an SE(3) flow model for LEGO brick reconstruction:
- Pretrained on procedurally generated brick assemblies, then finetuned with image conditioning;
- Combined with analytic connectivity-field guidance to infer part poses;
- The goal: reconstruct a physically buildable object structure directly from an image.
The project page includes visualizations going from input photos to full assembly sequences—an interesting intersection of visual understanding and physical assembly.
More from Embodied
- Musk: Tesla may begin December trials letting owners send their cars out as robotaxis — Polymarket · 2026-11-17
- Chart: Tesla's Austin Robotaxi Weekly Miles Hit 438K, Overtaking Waymo — ChrisGPT · 2026-11-17
- Remedy Robotics raises $62M led by DCVC to run its pivotal trial — hardimanjames · 2026-11-17
- Figure CEO: 1,500+ robots outnumber staff, Helix 3 is 'baby AGI' — PeterDiamandis · 2026-11-17
- Origami as physical reservoir computing: powering robotic crawling in Scientific Reports — cephaloform · 2026-11-17
- Robotics devs debate: superhuman dexter hands are here, whole-body control is the real bottleneck — carlosdponx · 2026-11-17