LeRobot 0.7 adds live robot steering — and breaks 0.6 setups

The open robotics stack adds remote inference and no-download dataset streaming, but forces a toolchain migration. Also: a 5.7× RL refit and a Codex skill.

Nowline Nov 7 9:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Live steering and remote inference land in LeRobot 0.7

    Run `lerobot-rollout --interactive=true` and the robot, weights, and processors stay loaded between commands: `/subtask` changes the instruction mid-run, `/vqa` asks what the robot sees, and `/autosteer` breaks a goal into subtasks. `--dataset.streaming=true` trains straight from the Hub or LanceDB with no local download, and `lerobot-policy-server` runs the policy on a GPU box over Zenoh+TLS while the robot handles sensors and motors. New model and robot support ships too, including Cosmos 3 and the Unitree G1.

  • Upgrading from 0.6 will break things — plan for it

    The jump isn't free. Weights & Biases is swapped for Trackio (`--wandb.*` becomes `--trackio.*`), FSDP1 is gone so 0.6.x FSDP checkpoints can't resume, Linux torch now ships as CUDA 13 and needs NVIDIA driver 580.65+, the `lerobot[async]` module is removed, and `--policy.use_peft=true` now trains LoRA instead of a full fine-tune. Seven image augmentations were silent no-ops on uint8 frames before this release, so retrain before comparing runs. Pin 0.6, install 0.7 in a fresh venv, and test in sim first.

  • NVIDIA cuts RL weight refit up to 5.7× in NeMo RL

    A new NCCL Reshard collective reframes the RL weight refit as a DTensor-to-DTensor reshard, so GPUs exchange only the shards they need instead of a full gather-broadcast. NVIDIA reports 2.7×–5.7× faster refits in NeMo RL benchmarks. The refit is pure dead time between a training step and the next rollout, so if you run RL post-training loops, that's wall-clock back on every iteration.

  • A Claude Code / Codex skill reverse-engineers any video

    rea-video is an Apache-2.0 agent skill for Claude Code and Codex that walks a clip shot by shot, tags each shot as live-action, AI-generated, or code/motion-graphics, and drafts a plan to recreate it. The author reports 98–99% shot-classification accuracy, measured on their own tuning clips — treat it as a starting point, not a guarantee. A weekend-sized way to turn a reference video into a build spec.