Glossary
Robot learning glossary
The words that show up in the client, in the training form and in every LeRobot thread, each defined in one sentence you can quote without reading the rest.
- 16 terms
- Written for the SO-100 workflow
- One page per term
This page is the whole vocabulary in one place. Every entry has a definition that stands on its own, so most of the time you will not need to open anything. Where a term carries more than a sentence of consequence, for example because it decides whether a trainer accepts your data, the linked page explains what follows from it.
A
- Action chunkingalso action chunk, chunk size, chunkSize, nActionSteps, receding horizon execution
- Action chunking means predicting a block of future actions from a single observation and executing part of that block before looking again. ACT is named after it: chunkSize is how many actions the model predicts, nActionSteps is how many of them the runtime executes, and the lerobot default is 100 for both.
C
- Calibrationalso calibrate, zero point, joint calibration, homing
- Calibration teaches the software where each joint of your particular arm physically sits, by capturing a zero reference in a known neutral pose and then sweeping each joint to learn its usable range. Every recorded action and every trained policy is expressed relative to that reference, so an uncalibrated arm produces data nobody can use, including you.
- Checkpointalso model checkpoint, saved weights, ckpt, saveSteps
- A checkpoint is the model’s weights written to storage at a particular training step, so one run leaves several of them behind rather than a single final model. The last checkpoint is not automatically the best one: which one wins is settled by attempts on the arm, not by the loss curve.
D
- Degrees of freedom (DoF)also DoF, DOF, degrees of freedom, 5 DoF robot arm, 6 axis arm
- Degrees of freedom counts the independently controllable joints of a robot. The SO-100, SO-101 and Koch v1.1 each have five joints plus a gripper, six actuators in total, which is one short of what an arm needs to place its tool at an arbitrary position and an arbitrary orientation at the same time.
E
- End-effectoralso gripper, end effector, EEF, tool center point, TCP
- The end-effector is whatever sits at the end of the kinematic chain and touches the world. On the SO-100 family it is a parallel-jaw gripper driven by the sixth servo, which makes it one more dimension of the same action vector rather than a separate device with its own interface.
- Episodealso demonstration, demo, trajectory, recording
- An episode is one recorded demonstration of a task, from the scene reset to the moment you stop, holding the synchronized joint states, actions and camera frames plus a single task sentence. Datasets are counted in episodes, and the episode is the unit you keep or throw away.
F
- Fine-tuningalso finetuning, fine tuning, transfer learning, adapting a pretrained model
- Fine-tuning starts from a checkpoint that already learned from other robots’ data and adapts it to your dataset, instead of starting from random weights. Four of the five models here are fine-tuned. ACT is the one trained from scratch, which is why it runs by far the longest schedule.
- Flow matchingalso flow matching policy, flow-matching action head, pi0 flow matching, continuous action generation
- Flow matching generates an action chunk by starting from noise and carrying it toward a valid trajectory along a learned velocity field, over several refinement passes instead of one forward pass. Pi0.5 is the model here that uses it, which is why its commands are continuous values rather than picks from a fixed set of bins.
I
- Imitation learningalso behavior cloning, behaviour cloning, learning from demonstration, LfD
- Imitation learning trains a robot policy from recorded demonstrations of a task instead of from a reward signal the robot optimizes by trial and error. The network learns the mapping from what the cameras and joints observed to the commands you issued, which means your demonstrations are not just training material, they are the specification.
- Inference latencyalso inference time, latency, ms per action step, policy control rate
- Inference latency is the wall-clock time a policy needs to produce one action step, measured in the cloud pool for each model here. It runs from about 20 ms for ACT to about 485 ms for Pi0.5, which is the difference between an arm that moves continuously and one that visibly pauses between decisions.
L
- Leader-follower teleoperationalso leader arm, follower arm, leader follower, master-slave teleoperation, puppeteering
- A leader-follower setup uses two arms of the same design: you move the leader by hand and the follower mirrors its joint angles in real time. The follower is the arm that gets recorded, so its calibration and its serial port are the ones that decide whether a dataset is usable.
- LeRobot datasetalso LeRobot format, LeRobotDataset, lerobot v2.1, lerobot v3.0
- A LeRobot dataset is the on-disk format this platform records into: JSON metadata under meta/, one Parquet file per episode for the numeric streams, and encoded video per camera. It is the common interchange format for low-cost manipulation data, and its version matters, because the GR00T trainers accept v2.1 only.
P
- Policyalso robot policy, visuomotor policy, control policy, trained model
- A policy is the trained function that maps an observation, meaning camera frames plus joint state and for the language models a task sentence, to the actions the arm executes next. On this platform the word usually means the concrete artifact you trained and named, not the algorithm class it belongs to.
T
- Teleoperationalso teleop, tele-operation, remote robot control, robot arm teleoperation
- Teleoperation is a human driving the robot in real time instead of a policy driving it. On this platform it is how every demonstration gets recorded: from the browser cockpit, from the desktop client, or by moving a second physical arm that the robot mirrors joint for joint.
- Training stepalso step, max steps, maxSteps, training iteration, optimizer step
- A training step is one optimizer update: draw a batch, compute the loss, adjust the weights. Every trainer on this platform counts steps rather than epochs, so the amount of learning in a run is fixed by the schedule and does not grow when your dataset does.
V
- Vision-language-action model (VLA)also VLA, VLA model, vision language action model, vision-language-action policy
- A vision-language-action model takes camera frames, the arm’s joint state and a natural-language task sentence as input and outputs joint targets for the next steps. Four of the five models trained on this platform are VLAs. ACT is the exception, because it reads no language at all.
Where these words are used
A definition is rarely the thing you came for. These are the pages where the term turns into a decision or a command you have to type.
Tutorials
From an unpacked kit to a trained policy driving the arm, in four steps.
Policy models
What each model costs to train, how fast it runs, and how much data it needs.
Training guides
One guide per model and arm, with the defaults the backend really sends.
Troubleshooting
When the term you just looked up appears in an error message.
Easier to understand with an arm in front of you
A real SO-100 is online and free to drive from the browser. Teleoperation, joint limits and the gripper stop being abstract after a minute of it.