LeRobot 0.7.0 upgrade guide: checkpoints, logging and licenses

Hugging Face released LeRobot 0.7.0 on November 5, 2026, expanding the robotics library with interactive language control, remote inference and a rebuilt training stack. For an existing project, the first question is whether the upgrade can preserve your run, not which new model has the highest score.

The most consequential changes sit below the demos: older FSDP training checkpoints cannot resume in the new layout, experiment logging moves to Trackio, and the previous asynchronous inference module is removed. This guide separates those migration decisions from optional new capabilities. It is based on the release, versioned documentation and model licenses, not hands-on robot testing.

Choose the migration path before installing

Keep a copy of the working environment, launch command, dataset revision and checkpoint outside the new environment. Then identify the kind of continuity you need. Reloading trained policy weights, resuming optimizer state and reproducing the same sample sequence are three different acceptance criteria. A successful model load proves only the first.

An unfinished FSDP run needs special treatment

The 0.7.0 training guide says FSDP checkpoints written by 0.6.x and earlier cannot be resumed in 0.7.0’s new training path because their optimizer-state layout differs. Finish an important old run in its pinned environment, or deliberately start a new run from suitable saved policy weights. Do not label the latter an uninterrupted resume.

The replacement uses FSDP2 and HSDP. Accelerate remains a launcher, but its YAML engine configuration is unsupported by the normal path. Use LeRobot’s distributed-configuration flags instead.

New checkpoint formats also serve different jobs. The default safetensors format provides directly loadable weights. DCP stores shards for distributed save and resume. The safetensors_dcp format keeps both. A DCP-only policy must be converted before ordinary policy-path loading. Confirm the exact artifact your deployment expects before choosing the cheaper training save.

Effective batch size combines micro-batches, workers and gradient accumulation. LeRobot does not automatically adjust learning rate or steps, so scaling to more GPUs can change the experiment.

Trackio changes where experiment logs can appear

The installation and tracking migration guide maps wandb.enable and wandb.project to trackio.enable and trackio.project. Local logging needs no account. Hosted logging uses a Hugging Face Space selected with trackio.space_id.

Check visibility before sending a real run. A newly created Space is public by default unless its organization has a private default. The trackio.private=true setting requests private creation. Run records include the launch command and console output, including tracebacks. The capture_console=false setting excludes console capture, but should not be treated as a universal secret scrubber.

For a local-only trial, omit space_id and check that TRACKIO_SPACE_ID and TRACKIO_SERVER_URL are unset. A command that looks local can otherwise inherit a hosted destination. Review a short synthetic run before logging private dataset names, internal paths or proprietary experiment output.

The same installation guide requires Python 3.12 or newer. Match the policy’s dependency instructions to the platform rather than assuming the broadest optional install is the safest choice.

Streaming saves downloads, but exact resume has conditions

The new reader fetches the Parquet data and MP4 byte ranges needed for training instead of first downloading the full dataset. That changes the storage bottleneck, not the need to budget memory and network capacity.

The streaming documentation distinguishes deterministic frame ordering from bitwise replay. Sample-exact resume requires the same world size, batch size and worker count. Checkpoints from the earlier multi-worker sampler cannot reliably be recognized for that guarantee. The documented alternative is a new run from their policy weights. Random image augmentations can still differ across runs.

The configured byte budget covers compressed payloads, not total process memory. Decoders, decoded tensors and other buffers add to it. A run can also spend time building an MP4 index sidecar when no valid cached or published index exists. Keep the dataset fixed during the run and measure startup time, steady-state throughput and peak memory separately. Training reads these sources without publishing sidecars or modifying the dataset repository.

Live steering does not always mean uninterrupted movement

An interactive session keeps the connection and policy available between instructions. The rollout documentation distinguishes three controls: /subtask changes an instruction, /vqa asks about the current observation, and /autosteer delegates subtask choice toward a goal. VQA and autosteering need a compatible text-generating policy or an external vision-language planner.

There is an important difference between those two routes. The policy’s own text generation can pause asynchronous motion, discard queued actions and resume from a fresh prediction. An external planner answers on a worker thread while the robot acts, although starting autosteer pauses the backend until the first planning reply. Instruction changes can also enter a hold.

Evaluate the transition, not just the answer. A correct description of a cube does not show that the robot retained progress after replanning. Test a command change during a simple task and record whether the resulting action remains appropriate. These controls are not evidence of general autonomy or a guarantee that motion will be continuously blended.

Remote inference needs a compatible robot and a timing budget

The removed async extra and lerobot.async_inference module are replaced by the remote extra, lerobot-policy-server and the remote rollout engine. Cameras and motor control stay with the robot while a separate machine runs the policy.

The remote guide requires a compatible action-chunk contract and a driver that supports position holding. Software support and hardware validation are separate entries in its support tables. Do not assume every model and robot combination is validated merely because it imports.

Old chunk thresholds and aggregation weights do not translate directly into the new refill, alignment and blending settings. Start from one working synchronous task and measure request turnaround against usable buffered playback. More generous timeouts cannot create missing action predictions.

Also test the documented failure path in simulation or a controlled setup: connection loss does not automatically restore the session when the network returns. Holding a target does not turn torque off or guarantee physical immobility. Follow the hardware’s safety procedures and preserve an independent operator stop.

Three model-license checks worth doing before a pilot

The release announcement adds several world action models and vision-language-action policies. Integration into one library does not make their weights, code or commercial rights interchangeable. These examples illustrate why the exact checkpoint matters.

G0.5: noncommercial code and weights

The G0.5 integration guide identifies both its policy code and weights as covered by the noncommercial G0.5 Community License, rather than LeRobot’s general Apache license. Redistributing a fine-tune also requires the relevant LICENSE and NOTICE files and identification of changes. LeRobot does not copy those notices automatically.

FLUX 3 Action: conditional commercial permission

The FLUX Kommunity v1.0 license allows qualifying users to generate and commercially use outputs, including action predictions, subject to its conditions. Its qualifying-user definition requires gross annualized revenue below $5 million across the user and affiliates. Larger commercial users need separate permission. Calling the model either unrestricted open source or simply noncommercial would miss that distinction.

The LeRobot FLUX guide supports inference and LoRA adaptation, with action and conditioning heads trained fully. Full fine-tuning and new embodiments use the upstream flux-action workflow. NATTEN compatibility can require a different supported PyTorch version. Check access to shared encoders as well as the robot checkpoint. The docs describe authentication if a repository is restricted, not a universal access gate.

Cosmos3: a base model is not a trained robot policy

The converted Cosmos3-Nano card identifies OpenMDW 1.1. NVIDIA’s model card lists commercial and noncommercial use. The integration guide distinguishes base checkpoints, whose action modules need robot training, from task-trained policies such as the LIBERO variants. Downloadability does not establish readiness for your arm, cameras or action space.

For context, our Genie 4 coverage examines a different world-model application: generating an interactive scene. In LeRobot, the decisive handoff is to a trained robot action policy. A convincing generated future and a successful physical action need different evidence.

A practical acceptance test for an existing project

Our recommendation is to change one layer at a time. Use a small, representative task and keep the old environment available until the new path passes the checks that matter to your project:

• Load the exact policy and processors with the expected camera names, action ordering and units before adding remote execution.
• Run a short training save-and-resume trial. Record whether you are testing weight loading, optimizer continuity or exact sample order.
• Inspect the tracking destination and a sample log before enabling hosted logging for real work.
• Compare a synchronous task with remote operation, then introduce instruction changes. Record completion, pauses and recovery separately.
• Review the checkpoint license and redistribution notices before a commercial pilot or sharing a fine-tune.

This sequence turns a broad release into a diagnosable upgrade. If the new model loads but a run cannot resume, or language steering works while task completion falls, you will know which decision needs attention. LeRobot 0.7.0 creates useful new options, but each project still needs its own evidence for continuity, control and permission.