🌐 The Big Picture: From LLMs to Robots That Act

Every embodied agent runs the same loop: perceive the world through sensors, reason about the objective and context, and act issuing motor commands. A pure LLM lives entirely in the middle layer—it operates on tokens, not the world. It has no sensory grounding (it doesn’t know where the cup is in metric coordinates) and no action channel (it can’t output torque). The VLA leap closes this loop: attach visual perception at the input and an action head at the output of the same backbone that already reasons.

O Moravec’s paradox (1980s) observes a counterintuitive reversal: tasks humans consider “easy”—picking up an object, maintaining balance, perceiving depth—require enormous sensorimotor computation, while “difficult” reasoning such as chess or integrals is relatively cheap for machines. The reason is evolutionary: billions of years refined our motor skills, making them unconscious; formal mathematics is only millennia old. That’s why a model can beat humans on math tests but still fail at folding a towel.

Classical robotics trained one policy per task per robot — N×M fragile models. The foundation model thesis inverts this: train one large model across many tasks and many bodies and harvest generalization. The bet is that diverse data induces shared representations — just as an LLM learns the syntax of one language and transfers it to another, a VLA learns to "grasp" with one arm and transfers that skill to another.

# Paradigma antigo            # Paradigma foundation
politica_pegar_braçoA()       modelo_vla(obs, instrução)
politica_empurrar_braçoA()      → ação  # qualquer tarefa
politica_pegar_braçoB()         # qualquer corpo
...  # N×M políticas frágeis   # 1 modelo, generaliza

RT-1 (Google, 2022) proved that a transformer could map camera + instruction to tokenized actions, trained on ~130k demos. But its knowledge stopped at the demos. RT-2 (2023) made the leap: it took a VLM already pretrained on the web (PaLI-X / PaLM-E) and co-fine-tuned it with robot data, representing actions as text tokens. The result was emergent semantic generalization — the robot could "pick up the extinct animal" (the toy dinosaur) by inheriting concepts from the web, without ever having seen that instruction in demos.

The field has organized around a few hubs. Google DeepMind kick-started academic research (RT-1/RT-2/RT-X). Physical Intelligence delivered the π0/π0.5 generalists with flow matching. NVIDIA pushes the humanoid stack forward with GR00T and simulation. Figure (Helix) and Tesla (Optimus) industrialize humanoid robots.

VLAs didn’t emerge from a single invention, but from the convergence of three curves. Data: Open X-Embodiment brought together 60+ datasets, 22 embodiments, and 1M+ trajectories. Compute: GPUs made it feasible to train models with 7B+ parameters on robotic video. Architecture: Transformers and diffusion/flow models provided the right form for ingesting multimodality and generating expressive actions.