build!: impose more strict constraint on rerun-sdk

[cverrier]

What this does

Originally, the constraint rerun-sdk>=0.21.0 was imposed. However, in rerun-sdk==0.23.0, Scalar was deprecated in favor of the newly introduced Scalars. For relatively new users of LeRobot that install a new virtual environment, rerun-sdk>=0.23.0 would be installed, hence raising multiple DeprecationWarnings when running robot experiments, which is annoying and clutters the logs. For example, running

HF_USER=$(huggingface-cli whoami | head -n 1)
python -m lerobot.record \
    --robot.type=so101_follower \
    --robot.port=<FOLLOWER_PORT> \
    --robot.id=follower \
    --teleop.type=so101_leader \
    --teleop.port=<LEADER_PORT> \
    --teleop.id=leader \
    --display_data=true \
    --dataset.repo_id=${HF_USER}/record-a-dummy-task \
    --dataset.num_episodes=1 \
    --dataset.push_to_hub=False \
    --dataset.single_task="A dummy task"

would lead to something similar to

/Users/clementverrier/repos/lerobot/lerobot/record.py:227: DeprecationWarning: since 0.23.0: Use `Scalars` instead.
  rr.log(f"observation.{obs}", rr.Scalar(val))
/Users/clementverrier/repos/lerobot/lerobot/record.py:232: DeprecationWarning: since 0.23.0: Use `Scalars` instead.
  rr.log(f"action.{act}", rr.Scalar(val))

for each recorded episode.

Therefore, replace Scalar with Scalars and update the associated dependency constraint accordingly (Scalars is only available in rerun-sdk>=0.23.0).

BREAKING CHANGE: requires rerun-sdk>=0.23.0 and replaces Scalar with Scalars, which is only available in rerun-sdk>=0.23.0. Users must upgrade their rerun-sdk version.

How it was tested

Run the same command as before:

```bash
HF_USER=$(huggingface-cli whoami | head -n 1)
python -m lerobot.record \
    --robot.type=so101_follower \
    --robot.port=<FOLLOWER_PORT> \
    --robot.id=follower \
    --teleop.type=so101_leader \
    --teleop.port=<LEADER_PORT> \
    --teleop.id=leader \
    --display_data=true \
    --dataset.repo_id=${HF_USER}/record-a-dummy-task \
    --dataset.num_episodes=1 \
    --dataset.push_to_hub=False \
    --dataset.single_task="A dummy task"

No DeprecationWarnings should be thrown anymore.

How to checkout & try? (for the reviewer)

See explanations above.

Since this dependency update introduces a breaking change (for any user that has rerun-sdk<0.23.0 in his environment`), I am not sure you want to merge this pull request directly. But I found it nice to have it as reference for future dependency update. @CarolinePascal could you look at it?

[Copilot]

Pull Request Overview

Updates the rerun-sdk requirement to v0.23.0+ and replaces deprecated Scalar calls with the new Scalars API to eliminate deprecation warnings.

• Bump rerun-sdk constraint in pyproject.toml
• Replace rr.Scalar with rr.Scalars in float logging within teleoperate and record loops

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
pyproject.toml	Bumped rerun-sdk version requirement to >=0.23.0
lerobot/teleoperate.py	Swapped rr.Scalar for rr.Scalars when logging floats
lerobot/record.py	Swapped rr.Scalar for rr.Scalars when logging floats

Comments suppressed due to low confidence (2)

lerobot/teleoperate.py:87

• Consider adding a test to verify that float observations and actions are logged correctly using rr.Scalars, ensuring no regressions and that deprecation warnings are gone.

                    rr.log(f"observation_{obs}", rr.Scalars(val))

lerobot/record.py:227

• Add a unit or integration test for the record_loop to ensure rr.Scalars is used for floats and no deprecation warnings are emitted.

                    rr.log(f"observation.{obs}", rr.Scalars(val))

[CarolinePascal]

Hi @cverrier,

Thanks for the PR ! Is rr.Scalar the only deprecation warning/deprecated use we should worry about ? I'm wondering if we should think ahead and start modifying our rerun usage to already match newly introduced usages.

If that is all, I'll give it a try on my side and proceed to merge :)

Best,

Caroline.

[cverrier]

Is rr.Scalar the only deprecation warning/deprecated use we should worry about ? I'm wondering if we should think ahead and start modifying our rerun usage to already match newly introduced usages.

I don't know, I just saw the warnings that I mentioned earlier and decided to fix them, I didn't try to modify anything else 😄

[ben-z]

Garbage collection (and rr.log(..., static=True)) is broken in rerun 0.23. I was wondering why rerun always uses over 10GB of RAM in my setup.

Comparison with LEROBOT_RERUN_MEMORY_LIMIT=100MiB:

export RUST_LOG=debug LEROBOT_RERUN_MEMORY_LIMIT=100MiB; python -m lerobot.teleoperate --robot.type=so101_follower --robot.port=$F1_PORT --robot.id=f1 --robot.cameras="${CAMERA_CONFIG}" --teleop.type=so101_leader --teleop.port=$L1_PORT --teleop.id=l1 --display_data=true

v0.22.1

v0.23.0

I recommend always using a max version in requirements.txt. I usually set it to the next minor version, e.g.

rerun-sdk>=0.22.0,<0.23.0

Made a PR for pinning the ceiling versions: #1435

[CarolinePascal]

Hi @cverrier @ben-z !

Recent changes on pyproject.toml enforced <0.23.0 on rerun 👍 We will try to bump directly to version 0.24.0 which will require extra work to support the new URDF loader. Feel free to keep discussing the topic on #1375/#1544 ! In the meantime, I'll close this PR.

Best,

Caroline.