LLM2R recasts 3D part assembly as sequence modeling with a decoder-only LLM
NationalUniversityofSingapore · hf · 2026-10-21
Researchers from NUS and PolyU present LLM2R, which treats 3D part assembly as sequence modeling: a Point-BERT encoder and Q-Former compress each part into geometric tokens, and a LoRA-adapted causal decoder-only LLM predicts all SE(3) part poses in a single forward pass. Learnable summary tokens after each part block let hidden states aggregate the whole assembly at constant token cost, and a part-aware pose head fuses each part's states for rotation and translation. One architecture handles semantic and geometric assembly, improving scalability over pairwise-relation or iterative generate-and-verify methods.
More from Embodied
- Mercedes, Volvo, VW, JLR join NVIDIA Hyperion network for robotaxis, Huang announces — testingcatalog · 2026-10-21
- Skills will generalize: the robot that cooks your breakfast will do this too — Dr_Singularity · 2026-10-21
- Motubrain 2.0 live-streams its full Physical AI training pipeline for two weeks — vllm_project · 2026-10-21
- Snap's $2,195 AI glasses tap Robert Downey Jr. to avoid another 'Glassholes' moment — aakashgupta · 2026-10-21
- Unitree Unveils H2-X "Kung Fu Legend" Robot With Marked Durability Gains — PsychologicalSoup251 · 2026-10-21
- Unitree H2 humanoid demos blocking knife attacks and lifting cars in rescue scenarios — CyberRobooo · 2026-10-21