RoboBoom
TNG Technology Consulting GmbH· June 3, 2026· 51:16

Big Techday 26: LeRobot: Making AI robotics more accessible with end-to-end learning - Steven Palma

Steven Palma, LeRobot tech lead at Hugging Face, explains how LeRobot makes end-to-end robotics learning more accessible at Big Techday 2026 in Munich.

In this talk

  1. 4:18

    A shared toolkit for robotics learning

    Palma presents LeRobot as an open source project that brings a common workflow to robot learning, much as Hugging Face’s Transformers library unified parts of natural language processing. Its six stages cover hardware, teleoperation, data collection, model training, real-world inference, and benchmarking.

  2. 11:24

    Low-cost robots and shared data

    The SO-100 arm costs about $100 to $120, and users can replace broken parts or print new ones. LeRobot standardizes datasets so teams can share demonstrations through the Hugging Face Hub; Palma says the project had about 100 terabytes of open data, mostly video and images plus robot state, with only around 15% from SO-100 platforms.

  3. 19:08

    Models, evaluation, and iteration

    LeRobot supports policies including ACT and the GR00T 1.5 vision-language-action model, alongside reward models that can help identify useful demonstrations. Palma emphasizes that imitation learning depends on demonstration quality, and describes simulation as especially useful for evaluation, though it can also supply training data.

  4. 31:04

    Scaling from simple tasks to T-shirt folding

    Palma describes a T-shirt-folding experiment using a more capable robot, additional compute, and roughly 150 hours of physical demonstrations; it used three RGB cameras and motor encoder readings, with no simulated training data. He says the open hardware project also includes a biped robot costing about $2,500, with design files, electronics plans, software, and a training guide shared for others to build on.

  5. 29:34

    Open development and community participation

    A LeRobot hackathon brought together more than 2,000 developers, who collected and shared datasets and models. Palma encourages developers to try the Python library with an SO-100, and says the project is designed to accommodate different hardware, policies, and benchmarks.

Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.