Open-source real-world RL toolkit for LeRobot101 (SO101), with continuous model, algorithm, and dataset releases.
LeRobot101 is live now. AgileX PiPER robot-arm support is coming soon.

Project Website • Reproduce Current Release • Community Program • WeChat Draft (ZH)

Evo-RL is a continuous open-source program, not a one-time demo release.

• Current platform: LeRobot101 / SO101.
• Next platform: AgileX PiPER robot arms (coming soon, https://www.agilex.ai/).
• Goal: build a community where tasks on both platforms can be uploaded, reproduced, and benchmarked with shared protocols.

1. To our knowledge (within current open-source LeRobot ecosystem), Evo-RL is the first project continuously pushing real-world RL releases on LeRobot101.
2. End-to-end loop already implemented: data collection -> value -> indicator -> policy -> real-world re-collection.
3. Engineering-first design: intervention/outcome labels, dataset quality report, indicator-conditioned training integration, and reproducible CLI chain.
4. Community-first roadmap: open task uploads and cross-platform benchmark evolution.

As of commit 852b23cb (2026-02-26), compared to main, core RL tooling includes:

• 101 core files changed (website and outreach files excluded)
• +6336 / -3224 lines
• New CLIs:
  • lerobot-human-inloop-record
  • lerobot-dataset-report
  • lerobot-value-train
  • lerobot-value-infer
• Indicator-conditioned policy path integrated into lerobot-train
• Value/advantage/indicator write-back pipeline for iterative real-world training

Recompute stats with:

git diff --shortstat main..kye/main -- . \
  ':(exclude)website' \
  ':(exclude)docs/outreach' \
  ':(exclude).github/workflows/deploy-website-pages.yml' \
  ':(exclude)README.md'

[HIL collection on real robot]
         |
         v
[Dataset quality report]
         |
         v
[Value training]
         |
         v
[Value / Advantage / Indicator annotation]
         |
         v
[Indicator-conditioned policy training]
         |
         v
[Real-world rollout + next iteration]

• Ubuntu 22.04
• Python 3.10
• CUDA 12.x class environment
• SO-series leader/follower arm + USB camera (640x480@30)

git clone https://github.com/Elvin-yk/evo-lerobot.git
cd evo-lerobot
conda activate lerobot
pip install -e .
pip install -e ".[pi]"      # value pipeline deps
pip install -e ".[feetech]" # SO101 related deps

export DATASET_REPO_ID=<your_dataset_repo_id>
export DATASET_ROOT=~/.cache/huggingface/lerobot

lerobot-find-port
lerobot-find-cameras

Confirm:

• Correct follower/leader serial ports
• Correct camera index/path
• Current user has serial device permission

lerobot-human-inloop-record \
  --robot.type=so100_follower \
  --robot.port=/dev/ttyACM0 \
  --robot.cameras="{front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
  --teleop.type=so100_leader \
  --teleop.port=/dev/ttyACM1 \
  --dataset.repo_id=${DATASET_REPO_ID} \
  --dataset.root=${DATASET_ROOT} \
  --dataset.single_task="Pick and place the red block" \
  --dataset.num_episodes=20 \
  --dataset.push_to_hub=false

Hotkeys:

• i: toggle intervention
• s: mark success and end episode
• f: mark failure and end episode

lerobot-dataset-report --dataset ${DATASET_REPO_ID} --root ${DATASET_ROOT}
lerobot-dataset-report --dataset ${DATASET_REPO_ID} --root ${DATASET_ROOT} --json

Expected output:

• Terminal report with success/failure/intervention metrics
• JSON report for logging and experiment cards

lerobot-value-train \
  --dataset.repo_id=${DATASET_REPO_ID} \
  --dataset.root=${DATASET_ROOT} \
  --dataset.download_videos=true \
  --value.type=pistar06 \
  --batch_size=16 \
  --steps=2000 \
  --output_dir=outputs/value_train/evo_value_demo \
  --job_name=evo_value_demo

Expected output:

• Value checkpoints under outputs/value_train/evo_value_demo

lerobot-value-infer \
  --dataset.repo_id=${DATASET_REPO_ID} \
  --dataset.root=${DATASET_ROOT} \
  --inference.checkpoint_path=outputs/value_train/evo_value_demo \
  --inference.checkpoint_ref=last \
  --runtime.batch_size=64 \
  --acp.enable=true

Expected output:

• Dataset fields written/updated:
  • complementary_info.value
  • complementary_info.advantage
  • complementary_info.acp_indicator

Optional visualization export:

lerobot-value-infer \
  --dataset.repo_id=${DATASET_REPO_ID} \
  --dataset.root=${DATASET_ROOT} \
  --inference.checkpoint_path=outputs/value_train/evo_value_demo \
  --inference.checkpoint_ref=last \
  --acp.enable=true \
  --viz.enable=true \
  --viz.episodes=all \
  --viz.output_dir=outputs/value_infer/viz

CLI flags use --acp.*, where acp_indicator is the binary advantage-indicator field.

lerobot-train \
  --dataset.repo_id=${DATASET_REPO_ID} \
  --dataset.root=${DATASET_ROOT} \
  --policy.type=pi05 \
  --policy.pretrained_path=lerobot/pi05_base \
  --acp.enable=true \
  --acp.indicator_field=complementary_info.acp_indicator \
  --steps=3000

• Platform: dual-arm SO101 task iteration
• D0 dataset scale: 300 episodes, 413,134 frames, ~3.82h at 30 FPS
• Internal observed trend: 100% data + Full FT forms a viable baseline in current setup

More benchmark cards and protocol details are published on the project website as releases progress.

We are opening a community track for tasks on:

• LeRobot101 (SO101)
• AgileX PiPER robot arms (coming soon, https://www.agilex.ai/)

Current submission flow:

1. Open an Issue with task setup and success definition.
2. Share dataset link + exact training/inference commands.
3. Submit result table + short video for benchmark integration.

Project page is in website/:

cd website
python -m http.server 8000

• This repository is LeRobot-derived and keeps the lerobot-* CLI namespace.
• Evo-RL extends upstream with real-world RL loop tooling and continuous release workflow.

@misc{evorl2026,
  title        = {Evo-RL: Continuous Open-Source Real-World RL on LeRobot101 and Beyond},
  author       = {Evo-RL Contributors},
  year         = {2026},
  howpublished = {\url{https://github.com/Elvin-yk/evo-lerobot}}
}

Apache-2.0. See LICENSE.