Roboticz Lab ES
‹ News

LeRobot 0.7 adds interactive control and streaming datasets | AI robotics

LeRobot 0.7.0 can change subtasks during a rollout, train from remote datasets and run the policy away from the robot.

  • LeRobot
  • robot learning
  • open source
  • datasets
  • robot control
  • AI
  • news
Generated image: macro detail of lab instrumentation, accompanying the entry LeRobot 0.7 adds interactive control and streaming datasets

What has actually changed in the last 24 hours? LeRobot 0.7.0 is now available. The useful change is not another closed model. Hugging Face’s open project can change instructions while a robot stays connected, train without downloading a whole dataset first and place heavy inference on another computer.

This is a software release, not a new robot or an independent performance test. The features and speed figures below come from the LeRobot team.

Correct a task without rebuilding the session

lerobot-rollout now has interactive sessions. The release documentation says the robot connection, policy weights and processors remain loaded between instructions. From a terminal, you can change a subtask, ask what the policy sees or give it a broader goal.

For a desktop arm, that has a practical effect. If the gripper misses a part, you do not need to restart the whole chain just to adjust the instruction. You can record the correction in the same session. That makes it easier to collect data with failures and recoveries, which is usually what a home setup lacks once it leaves the perfect demo.

Not every policy can answer questions or select text-based subtasks. LeRobot says those actions need either a text-generating policy or an external VLM planner. Check that before expecting a control model to hold a conversation.

Training no longer starts with a full download

Version 0.7 adds streaming for LeRobot datasets from the Hub or a Storage Bucket. The reader fetches the video ranges needed during training instead of requiring a complete local copy. The team reports 2.1 to 2.5 times more samples per second with four workers in its tests. That is Hugging Face’s result, not a promise for every network or model.

It also adds a route for storing datasets as Lance tables while retaining the original MP4 files. If you are preparing a small dataset, the immediate lesson is simpler: separate data storage from the training machine’s disk. You can test the pipeline before buying an SSD for a collection you are still fixing.

The policy can run away from the robot

With lerobot-policy-server, cameras, motors, teleoperation and recording stay next to the robot while the model runs on a GPU elsewhere on the network. LeRobot uses Zenoh and documents TLS and access-control configurations. The client receives action chunks and adjusts late arrivals to the remaining trajectory.

What should you do first? Build the flow in simulation and measure your actual network latency before attaching motors. The release includes environment servers for evaluation and simulation, so you do not need physical hardware to check the basic path. Start with that short test. It often shows whether Wi-Fi, rather than the model, is the real problem.

Sources
  1. LeRobot v0.7.0 technical release by Hugging Face
  2. LeRobot v0.7.0 signed GitHub release

Every claim above comes from here. If something has no source, it's not in.

Now go build it

Guides with the bill of materials and the logic explained.

Subscribe by RSS