The biggest progress of recent years hasn't been in mechanics, but in software. Today, robots learn using the same methods as large language models.
Three types of models
- LLM (large language model): understands language and breaks tasks down into steps.
- VLM (vision-language model): understands images and language together, such as “Pick up the red cup.”
- VLA (vision-language-action): generates joint movements directly from images and language. Examples include Figure Helix, NVIDIA GR00T and models from Physical Intelligence.
How robots learn
- Imitation learning: People demonstrate tasks, often via teleoperation.
- Reinforcement learning: Millions of attempts in simulation.
- Sim-to-real: What was learned in simulation is transferred to the real robot.
Computing power
AI increasingly runs directly on the robot. NVIDIA Jetson modules are widespread; the Jetson Thor generation is becoming the standard in many high-end humanoids in 2025/26. Performance is measured in TOPS – trillions of operations per second.
What this means for buyers
- Ask which model is used and whether it runs locally or in the cloud.
- Check whether you can train your own skills.
- Distinguish between a demo video and an autonomous production feature: teleoperation is not autonomy.