Real robot demonstrations
Scarce, high-value physical data.
Fourier GR1 · Unitree G1NVIDIA / TRAINING
Better data.
Better robots.
GR00T N1.7 learns from human video, real robot demonstrations and simulation.
~32K hReal + human data
~8K hSimulation data
Published training-data snapshot
Code · Apache 2.0
Weights · NVIDIA Open Model License
THE DATA PYRAMID
Synthetic trajectories
Isaac · Cosmos · MimicGen
GR00T-Mimic · GR00T-Dreams
Internet + human video
Object semantics and human egocentric experience.
From learning to doing.
Pick-and-place: a foundation for tabletop and industrial manipulation.
01
See & grasp
Identify the object and establish a stable hold.
02
Move & place
Guide the object to its target and release.
03
Stay balanced
G1 reference task: move an apple onto a plate with whole-body control.
Example tasks: Apple → plate / basket · Bottle → cabinet · Can or cup → drawer · Fruit → basket
Pick & Place
NVIDIA’s humanoid pick-and-place task is simple on purpose: a Unitree G1 stands at a table, sees an object such as an apple, picks it up, and sets it on a plate or in a basket. In the open GR00T dataset the robot also follows a short language command and chooses among a few fruits. A harder variant adds walking, go to the object, grasp it, then carry it to the target.
It matters because this is the basic unit of useful robot work. The same short sequence tests vision, grasping, arm control, and balance. NVIDIA uses it as the starter loop for Isaac GR00T: collect a few demonstrations, fine-tune a policy, score it in simulation, then run it on the real G1. If a humanoid cannot reliably pick and place, it is not ready for kitchens, warehouses, or homes.
Context & sources
The pyramid describes complementary data sources, not measured volumes. Task examples illustrate a broader pick-and-place family. Artwork is conceptual.