Hugging Face Releases LeRobot v0.7.0 Robotics Training Platform, Adding Interactive Control and Large-Scale Training

Hugging Face released LeRobot v0.7.0, a robotics training platform, on November 5, 2026. The LeRobot Hub contains about 300TB of open physical AI data in LeRobot format, which can be used for training without first downloading it. The update also supports changing subtasks for robots while they are operating, asking questions about visual information and autonomously selecting subtasks based on goals.

Hugging Face’s announcement brings data streaming for training, distributed training for large models and changes to instructions while robots are operating into a single LeRobot platform. For robotics developers, the burden of copying data to training infrastructure and the operational challenges of running large models across multiple GPUs are implementation issues alongside model selection.

The update centers on connecting a Hub that stores about 300TB of data, data loading without downloads, distributed training and instructions for robots while they are operating. Beyond adding models, it treats the burden of moving data to training infrastructure and the computing and memory constraints of large models as features of the same platform. Having public data does not necessarily mean it can be used for training; mechanisms for loading the required video segments from storage and the ability to resume distributed checkpoints are likely to be practical requirements. Bringing language into robot training also extends work from the previous version. According to Hugging Face, v0.6.0 enabled language annotations to be stored in datasets, while v0.7.0 extends that capability to training and interactive operation. The new interactive features let users change instructions without stopping a robot or ask what it is seeing. However, questions and autonomous goal execution require a policy that generates text or an external planner. The ability to control a robot in natural language should be distinguished from whether any given policy has that capability. On the computing side, Hugging Face says world action models have 5–20B parameters and use many video tokens per sample, which can exceed the memory capacity of standard data parallelism. The addition of FSDP2, HSDP and distributed checkpoints can be understood as a change to support not only scaling up models but also continuing and resuming training across multiple GPUs. The issues overlap with those discussed in NVIDIA Explains Full-Stack Design to Address AI Memory Constraints (November 5, 2026), but LeRobot’s announcement is not about a product that increases GPU memory capacity; it addresses the partitioning of training workloads and data loading. The setup for collecting and annotating data, training, running models on robots and recording episodes during operation places data and policies in one development pipeline. However, the available information does not say how recorded episodes are reviewed or processed before being used for retraining, or how operational safety is verified. Next to watch are the operating conditions of each model on physical robots, data quality management and communications requirements for remote inference.

Data can be loaded from the Hub, Storage Bucket and LanceDB. For streaming standard datasets, only the required video data segments are retrieved. Hugging Face says the number of samples per second increases by 2.1–2.5× when using four DataLoader workers. A feature that moves image transformations to the GPU increased ACT training speed by 2.8× on an H100 and 3.3× on an A100, the company said. The update adds FSDP2, HSDP and distributed checkpoints for large models. According to Hugging Face, when using 6B FastWAM on an 8×H200 node, distributed checkpoint saving was 3× faster, and peak memory when resuming fell from 143GB to 42GB per GPU. One of the added models, NVIDIA’s Cosmos 3, comes in two versions: Nano (16.5B parameters) and Edge (4.6B). Their average success rates on LIBERO are reported at 97.90% and 97.95%, respectively. The FineART dataset comprises more than 40k bimanual manipulation episodes, 1,718 hours and more than 533k subtask annotations, and totals 1.9TB in LeRobot format.