Human-Centric Intelligence in the Era of Foundation Models: A Survey

This survey reframes human-centric intelligence as a connected foundation-model discipline rather than a collection of isolated vision, motion, interaction, and robotics tasks. Its central taxonomy spans 6 levels: visual appearance, spatial geometry, kinematic dynamics, interaction modeling, world simulation, and embodied agency, progressing from observable subjects to dynamic actors and situated agents. Methodologically, it organizes systems into single-step mapping, autoregressive sequential factorization, iterative generation with diffusion or flow matching, and hybrid architectures that combine reasoning, generation, and control. It also compares training strategies including full fine-tuning, LoRA-based parameter-efficient tuning, instruction tuning, reward optimization, distillation, and inference methods such as retrieval augmentation and guided sampling. Representative scaling trends include Sapiens, trained on 300M human images, and Sapiens2, expanded to 1B images for dense pose, parsing, depth, and surface-normal prediction. In motion and embodiment, MotionGPT and LLaMo treat human movement as a language-compatible modality, while world models such as EgoSim connect action-conditioned prediction to planning. Human experience is increasingly used for robot learning: NVIDIA’s GR00T N1 uses 8.4K hours of data and a 2.2B-parameter policy, while GPT-4 exemplifies the role of general-purpose vision-language reasoning in humanoid control. The survey’s main conclusion is that scale alone is insufficient: future systems must align multimodal human representations with metric geometry, physical constraints, causal world states, privacy-aware data, and closed-loop execution, evaluated through integrated benchmarks rather than task-specific scores alone.