What Is a Humanoid Robot?

A humanoid robot is a robot whose body follows the human layout: a head carrying cameras, a torso, two arms with hands or grippers, and usually two legs. Some designs replace the legs with a wheeled base and keep a human-shaped upper body.

The reason for the shape is practical rather than cosmetic. Factories, warehouses, hospitals and homes are built around people, with stairs, door handles, shelves at human height and hand tools. A robot with a human-like reach and footprint can, in principle, work in those places without rebuilding them, and can learn from people by copying how they move.

Humanoid robots are not new. Waseda University’s WABOT-1 in 1973 is often cited as the first full-scale humanoid, and Honda’s ASIMO, shown in 2000, made walking robots famous. What has changed since around 2022 is the AI. Large models trained on video, simulation and human demonstrations now let humanoids learn general skills, and a wave of companies, including Agility Robotics, Boston Dynamics, Figure, Tesla, Unitree and Apptronik, are building them for work.

Humanoid design is a balance between strength, weight, battery life and dexterity. Each part of the body brings its own engineering problem.

Humanoids join robot types that already do much of the work in industry. Each suits a different job.

For a single repeated task, a fixed arm or a collaborative robot is usually simpler and faster. A humanoid makes most sense where the tasks change, the space cannot be rebuilt, or one robot must cover several jobs.

A humanoid runs several AI models at once, split by how fast they must react. At the top, a robot foundation model, often a vision-language-action model such as NVIDIA Isaac GR00T N1, interprets an instruction and the scene and decides what to do. Below it, learned motion policies turn that decision into walking, reaching and grasping. These policies are mostly trained with reinforcement learning in simulation and by copying human demonstrations captured through teleoperation or motion capture.

Beneath the AI sits a conventional robot controller that closes the joint and balance loops at high rates and triggers protective stops. Keeping that layer separate is what lets a learned model drive the robot without being trusted with its basic safety.

Most humanoid deployments are still trials, but they cluster in a few places where people currently do repetitive handling work.

The onboard computer has to fit in the torso, run on battery and still process many camera streams alongside large AI models. NVIDIA developed Jetson Thor with humanoid robots as a main target, quoting up to 2,070 FP4 TFLOPS and 128 GB of memory within a configurable 40 to 130 W range. Our Jetson Thor robot computers add up to 16 GMSL camera inputs, 10 GbE and CAN FD for the joint controllers, and our Jetson Thor and AGX Orin comparison explains what the extra compute means for humanoid workloads.

We supply the onboard AI computers for humanoid and mobile robot projects on NVIDIA Jetson Thor and Jetson Orin, with the cameras, storage and JetPack image fitted and validated as standard. Send us your sensor list, power budget and control interfaces, and we will specify the system to match.