Vision-language-action

VLA training data connects instructions, perception, and control

Structure a VLA dataset around task language, observations, action targets, embodiment, timing, and outcomes.

Synthetic schema sampleObservation → action → next state
01
00.000sObservergb.player_view
02
00.050sActmove.forward + camera.delta
03
00.100sTransitionpose.x +0.84 m
04
00.150sOutcomeprogress 0.31 → 0.36

The values are synthetic and illustrate the public specification. They are not measurements from a commercial dataset.

Direct answer

A VLA dataset pairs a language instruction and visual context with the action sequence an embodied agent should execute. Strong VLA datasets preserve action semantics, robot configuration, temporal alignment, task progress, and the difference between commanded and measured motion.

Language describes the task, not the whole trajectory

A task string can express intent while images describe the current scene. The action target still needs a declared interface: pose delta, joint target, end-effector command, gripper state, discrete skill, or another embodiment-specific control.

Paraphrases, corrections, subgoals, and termination language should remain linked to the episode segment they govern. Treating one caption as supervision for an entire session can blur which instruction caused which behavior.

  • Instruction, goal, and optional subgoal text
  • Camera and sensor observations with calibration
  • Action tokens plus physical units and control rate
  • Embodiment, task, success, failure, and intervention metadata

Action tokenization must remain reversible

A VLA may predict discretized tokens, continuous action chunks, or higher-level skills. The release should preserve the original command beside any quantized or normalized representation.

Token vocabulary, bin boundaries, chunk horizon, control frequency, coordinate frame, and clipping behavior belong in the schema. Otherwise identical token IDs can imply different motion across robots or datasets.

Cross-embodiment mixtures need explicit normalization

Large VLA programs often mix tasks and robot types. Positive transfer is possible, but morphology, sensors, action spaces, control rates, and success definitions do not become equivalent because they share one loader.

Keep source-dataset identity and embodiment metadata available during training and evaluation. Hold out robots, scenes, objects, and instruction families to measure generalization rather than memorization.

Questions this page should settle

What is a vision-language-action model?

A VLA model uses visual observations and language context to predict actions for an embodied agent, commonly a robot.

What fields belong in a VLA dataset?

Include language goals, time-aligned observations, original and model-ready actions, embodiment and calibration metadata, task state, interventions, and outcomes.

Is web-scale image and text data enough for a VLA?

It can provide visual and semantic priors, but action prediction still requires embodied action data or another grounded control signal.

Bring us the training target.

We will map it to the observations, actions, state, outcomes, rights, and delivery evidence the dataset must contain.

Request access