← Learn

Choose data for a manipulation experiment

Human motion, simulated trajectories and robot demonstrations answer different questions.

Write down what your learner needs

Does your experiment need images and poses, contact annotations, robot actions, or complete task trajectories? Specify the hand, sensors and task before choosing a dataset. A large collection can still lack the labels you need.

Match data to the question

ARCTIC and OakInk2 are starting points for bimanual human–object motion. DexMimicGen provides simulated robot demonstrations. DROID and Open X-Embodiment provide broader robot-learning data; do not assume they contain multi-finger hand actions.

Check one sample and its terms

Use the publisher’s access route and inspection tools. Check coordinate frames, units, action definitions, missing observations and train/test splits. Read dataset, code and body-model licenses separately. ARCTIC restricts commercial use and redistribution; DexMimicGen data and code have different licenses.

Make the choice explainable

Record the dataset release, task, inputs, labels, access terms and the main mismatch with your intended robot. Share that comparison and a small permitted result before investing in a larger training run.

ARCTIC project and access · OakInk2 project · DexMimicGen release and licenses