The robot AI glossary for newcomers

Robotics conversations assume vocabulary nobody stops to explain. Here are the 15 words that unlock most threads, in the order you will meet them.

The data words

Episode: one recorded attempt at a task, start to finish. Teleoperation: driving the robot by hand while everything records; puppeteering with a purpose. Task label: the sentence saying what an episode shows; placeholder labels like "task desc" quietly poison training. Camera stream: one camera's video in the dataset; names matter and must stay consistent. The LeRobot format: the standard way episodes, cameras, and labels get packaged so everyone's tools read everyone's data. The Hub: the public library where datasets and models live.

The model words

Policy: the trained brain file: cameras in, motor actions out, many times per second. VLA: a policy that also reads language, so you can tell it what to do. Fine-tuning: teaching a pretrained model your specific task with your episodes; cheap compared to training from nothing. ACT and SmolVLA: the two trainable-at-home model families most beginners meet first.

The reality words

Embodiment: which body the robot has; data from one body does not automatically fit another. Sim vs real: simulation is the video game version; useful, fast, and it lies a little, which is why real-hardware results matter. SO-101: the community's favorite cheap arm and the assumed default in most tutorials. Junk gate: the basic checks a dataset must pass to be worth anyone's time. Facet: one check's result; facts before grades.

Where these words come from

This glossary is the public version of the field notes we keep while building Datum. When a word confuses newcomers in community threads, it earns a place here. Missing one? Tell us: hello@datumrobotics.ai