Explainers

How humanoid robots learn: teleoperation, demonstrations and simulation

By Humanoid News Staff · Published 9 November 2026 · Last updated 9 November 2026

Short answer: Most of today's humanoid robots learn tasks from examples rather than hand-written code. People demonstrate the task, often by remotely operating the robot (teleoperation), and those recordings train an AI policy. Companies add simulated practice and synthetic data because real demonstrations are slow to collect. Once trained, a skill can be copied to every robot in a fleet.

Teleoperation data. Figure says its Helix model was trained on about 500 hours of teleoperated behaviour from multiple robots and operators, with a separate vision-language model writing instructions for each clip after the fact. Google DeepMind's RT-1 and RT-2 work used demonstrations that, according to DeepMind, were collected with 13 robots over 17 months. Physical Intelligence says pi0 was pre-trained on data from many different robots and then fine-tuned on high-quality data for hard tasks such as folding laundry.

Remote help in the home. 1X takes the same idea into customers' homes. Its NEO product page describes an Expert Mode in which, for tasks NEO does not know, an expert from 1X can remotely supervise the robot at scheduled times to help it learn. Owners can also pilot NEO remotely from the app or a VR device. More in our NEO explainer.

Simulation and synthetic data. NVIDIA's GR00T workflow lets developers capture demonstrations through teleoperation in simulation or on a physical robot, post-train GR00T models on real or simulated data, and test the result in its Isaac Lab-Arena simulator before deployment, according to NVIDIA's developer page. When NVIDIA launched GR00T N1 in March 2025 it said the fast action part of the model was trained on human demonstration data and a large amount of synthetic data.

Fleet learning. Boston Dynamics says that when one Atlas learns a new skill, the task can be deployed across a customer's whole Atlas fleet. That is a big difference from human training: a skill is learned once and copied. See electric Atlas explained.

Why it is still hard. Physical Intelligence notes that unlike language models, which learn from a large share of the documents on the web, robotics has no equivalent pool of data, so data for a new skill often has to be collected with a particular robot. Vision-language-action models try to reduce how much is needed; see what is a VLA model.

← All news