Every robot runs the same loop: perceive, decide, act, then start again, several times per second. The problem is that the physical world does not forgive mistakes.
2.1The fundamental loop
A robot perceives its environment through sensors, decides through computation, acts through actuators. Unlike an LLM's screen, the physical world is unstructured, changing and irreversible: a wrong word can be corrected, a broken glass cannot.
2.2The body: actuators, sensors, energy
- Actuators: electric motors and precision gearboxes, costly components dominated by a few specialist manufacturers.
- Sensors: cameras, lidar, inertial measurement units, force sensors, with one big absentee: fine touch.
- Energy: a current humanoid often lasts only a few hours on battery; the energy efficiency of living beings remains out of reach.
2.3The classic brain
Three disciplines keep a robot on its feet: control (correcting constantly), SLAM (localising while mapping) and motion planning (computing a collision-free trajectory). It all often comes together with ROS, an open-source middleware that has become the standard.
2.4The revolution: from programmed to learned
- Imitation: humans teleoperate the robot; the model learns from demonstrations.
- Reinforcement in simulation: the robot trains virtually before transferring to the real world.
- Foundation models: VLAs such as RT-2, π0, GR00T or Helix learn to link instruction, vision and action.
2.5Why it is still so hard
- Moravec's paradox: calculation is easy for a machine; picking up an object in a kitchen is very hard.
- The human form: useful in a world designed for us, but often less efficient than a specialised robot.
- The data wall: LLMs had the Internet; there is no Internet of physical movement.
- Hidden teleoperation: in front of a spectacular video, always ask who is really in control.
Key takeaway
A robot is three nested problems: a body, a brain and their encounter with an unpredictable world. Robotics is shifting from programmed to learned, but your hands remain a technology of formidable complexity.
Sources
H. Moravec, Mind Children; Google DeepMind, RT-2 and Open X-Embodiment; ROS documentation (ros.org); Nvidia Isaac Sim and GR00T.
Finished this module?
Your progress is saved on this device.