********Domain-specialized transformers and embodied/sim applications** [developing]
Key Questions
What hybrid transformer models are featured?
Hybrids include LISTA/scDynOmics, PINO for PDEs, PIXAL for physics, MMFormer for remote sensing, and GT-TD3 for high-DOF manipulators. FocusMamba achieves SOTA in brain tumor segmentation.
What embodied AI advancements are highlighted?
EgoSim for egocentric simulation, UniDriveVLA/CARLA-Air for driving, HandX with 54h bimanual data, and LeRobotHF for cloth folding with 100h+ real data. OpenWorldLib unifies World Models codebases.
How do World Action Models compare to VLAs?
World Action Models show better generalization than VLAs in robustness studies via GroundedPlanBench with 1k tasks. They enhance sim-to-real in robotics.
What is SMASH in humanoid applications?
SMASH enables humanoid ping-pong with onboard vision. It demonstrates embodied transformer use in dynamic physical tasks.
What medical and bio domain transformers exist?
MedOpenClaw, Medical AI Scientist, PhenoAssistant for plant phenotyping, and BioVITA/Lingshu-Cell for biomed. FocusMamba targets brain tumors, with neural-symbolic agents advancing.
What is EgoNav's capability?
EgoNav provides zero-shot humanoid navigation using Stanford's approaches. It supports embodied sim applications.
What benchmarks support domain transformers?
GroundedPlanBench (1k tasks), Colon-Bench for endoscopy, and NeurIPS Embodied Challenge test robustness. They evaluate VLAs and World Action Models.
What is BraiNCA?
BraiNCA introduces brain-inspired neural cellular automata for morphogenesis. It applies to embodied and domain-specific simulations.
Hybrids (LISTA/scDynOmics, PINO PDE, PIXAL physics, MMFormer remote sensing, GT-TD3 GNN-Transformer high-DOF manip, FocusMamba Mamba brain tumor SOTA, SciLT long-tailed scientific images, MedGemma 1.5 biomed); embodied (EgoSim, WorldCache/USV-3.0, V-JEPA, UniDriveVLA/CARLA-Air, HandX 54h bimanual, SoftMimicGen deformable sim data threading/whipping surgical, LeRobotHF cloth folding 100h+ bimanual real data/OpenWorldLib World Models codebase, GroundedPlanBench 1k tasks, Colon-Bench endoscopy, EgoNav zero-shot humanoid nav, SMASH humanoid ping-pong vision, World Action Models vs VLAs robustness, Action Images end-to-end policy multiview video gen); BioVITA/Lingshu-Cell bio, MedOpenClaw/Medical AI Scientist med + neural-symbolic biomed agents, PhenoAssistant plant phenotyping agents, Proteina-Complexa binders. Sim-data robotics sim-to-real VLAs; NeurIPS Embodied Challenge.