Skip to content

VLA

Vision-Language-Action (VLA) fine-tuning on MinT is Coming Soon. VLA models take images and language instructions as input and output robot actions; MinT will support OpenPI-compatible VLA training once the embodied base models land on mint.macaron.im and mintcn.macaron.xin.

For interest registration, contact sales@mindlab.ltd or Schedule a Demo and mention VLA in your request.

What this page will cover

When VLA lands, this page documents the same four-section shape as the other algorithm pages:

  • Configuration — mint.ServiceClient, create_lora_training_client for VLA-capable OpenPI base models, camera-layout selection, action-token budget.
  • Prompting Guide — multi-camera image placement, proprioceptive state vectors, and quantized action-token targets.
  • Output Format — how the model's action tokens are dequantized back to continuous control signals; per-step action ranges.
  • All Parameters — VLA-specific knobs (max frames, action dimensionality, quantization levels, autoregressive mask) layered on top of the standard SFT / RL parameters.

While VLA is in flight, the closest shipped pieces are:

  • VLM — vision-language model fine-tuning status.
  • SFT Overview — the supervised fine-tuning loop VLA reuses.