Dataset profile

CALVIN

CALVIN is an open-source simulated benchmark for language-conditioned robot manipulation. It is useful for evaluating long-horizon VLA-style policies, but it does not replace real healthcare task data.

Source institutionUniversity of Freiburg
Robot Manipulation

Overview

CALVIN is an open-source simulated benchmark for language-conditioned robot manipulation. It is useful for evaluating long-horizon VLA-style policies, but it does not replace real healthcare task data.

What's included

  • Simulated robot observations and actions
  • Language annotations
  • Static camera, gripper camera, depth, tactile, and proprioceptive options

Tasks / activities

open drawerpush blockstoggle switchslide doorlanguage-conditioned multi-step manipulation

Modalities

RGBDepthRobot ActionsLanguageTactile

Collection methodology

A simulated robot environment defines long-horizon manipulation tasks conditioned on natural-language instructions.

Potential applications

  • Language-conditioned control
  • Long-horizon manipulation
  • VLA policy benchmarking
  • Simulation-based pretraining

Access & licensing

This is a third-party public/research dataset. We do not own or distribute this dataset. Access and usage are governed by the original publisher's terms.

Access and use are governed by the CALVIN repository and dataset terms.

Limitations / opportunities for custom data

Teams validating policies in simulation may still need real care environments, expert demonstrations, and healthcare workflow labels.

Need data beyond this benchmark?

We design custom healthcare and caregiving collections around the tasks, experts, environments, and modalities your model requires.

Request Similar Data