Dynalearn aggregates six years of manipulation experiments, covering assembly, fastening and inspection with both rigid and soft grippers. Each sequence pairs low-level torque data with operator narrations transcribed into time-aligned tokens.

The release includes scripts to convert the dataset into Hugging Face datasets and ROS bag bundles. Research partners at ETH Zürich and UC Berkeley have already benchmarked multi-task transformers on Dynalearn, reporting faster convergence on contact-rich skills.

MIT licensed the dataset under Creative Commons for commercial use, provided companies share model evaluations back with the consortium. The lab is running an embodied AI challenge early next year to surface best-performing policy architectures.