Site Traffic

Visitors --
Views --

Continuous Open-Source Real-World RL

Evo-RL SO101 to AgileX PiPER

A LeRobot-aligned, full-pipeline real-world offline RL workflow for real robots: collect rollout data, train value functions and policies offline, infer advantage tags, and iterate deployment in a closed loop.

NEWS February 26, 2026 First SO101 real-world RL baseline + reproducible CLI released.

pipeline.sh

lerobot-human-inloop-record
  --dataset.repo_id=<DATASET>

lerobot-value-train
  --value.type=pistar06

lerobot-value-infer
  --acp.positive_ratio=0.3

lerobot-train
  --acp.enable=true

Robot Platforms

0

Open Task Datasets

0

Open Models

0

Release Year

0

SJTU community visual EvoMind

Project Focus

Open Infrastructure for Real-World RL

Open real-world RL pipeline

Runnable workflows on SO101 and AgileX PiPER covering collection, offline value/policy training, and iterative deployment.

Reproducible assets

Codebase, model layout, dataset structure, and command-line templates are released to reduce setup friction and improve reproducibility.

Community co-evolution

Reproduce existing real-world RL methods, extend value/policy variants, and continuously publish experiments and benchmarks for collective progress.

Method

6-Stage Closed-Loop Pipeline

  1. 01

    System preparation to ensure robot, sensors, and control stack are ready.

  2. 02

    Interactive data collection with human guidance and episode-level outcomes.

  3. 03

    Value modeling to estimate trajectory quality from accumulated data.

  4. 04

    Signal annotation to convert value outputs into usable training supervision.

  5. 05

    Policy update and deployment to run the improved policy in real tasks.

  6. 06

    Data refresh and iteration to continue improving robustness and success rate.

Results

Training Signals Across Iterations

Illustrative trend view for project webpage. Replace with your run logs for paper-ready plots.

What This Visual Captures

  • Value trend stabilizes as rollouts and labels improve across rounds.
  • Advantage separates better trajectories used for ACP tagging.
  • Success rate rises with iterative deployment and data refresh.

Current default ACP ratio in README pipeline: --acp.positive_ratio=0.3.

Qualitative Demos

Visual Results Gallery

VALUE VISUAL

Success and Failure Case

Frame-level value visualization on real robot rollouts.

Value visual success case
Value visual failure case

POLICY ROLLOUT

Deployment Visual Results

Final policy behavior clips in target task execution.

Policy rollout result 1
Policy rollout result 2

HUMAN-IN-THE-LOOP

Intervention and Recovery Results

Human guidance segments and autonomous recovery outcomes.

Human-in-the-loop result 1
Human-in-the-loop result 2

Architecture

LeRobot-Aligned VLA + Value Loop

VLA architecture

Evo-RL reuses LeRobot's practical data and inference stack while adding value-guided filtering signals (`value`, `advantage`, `acp_indicator`) to improve iterative policy training.

The same design supports single-GPU and accelerate-based multi-GPU workflows for both value and policy stages.

Reproducibility

Core CLI Commands

Human-in-loop Record

lerobot-human-inloop-record \
  --robot.type=<ROBOT_TYPE> \
  --teleop.type=<TELEOP_TYPE> \
  --dataset.repo_id=<HF_USER_OR_ORG/DATASET> \
  --dataset.single_task="<TASK_TEXT>" \
  --dataset.num_episodes=<NUM_EPISODES> \
  --display_data=true

Value Training

lerobot-value-train \
  --dataset.repo_id=<HF_USER_OR_ORG/DATASET> \
  --value.type=<VALUE_TYPE> \
  --batch_size=<BATCH_SIZE> \
  --output_dir=outputs/value_train/<RUN_NAME> \
  --job_name=<RUN_NAME>

Policy Training

lerobot-train \
  --dataset.repo_id=<HF_USER_OR_ORG/DATASET> \
  --policy.type=<POLICY_TYPE> \
  --steps=<TRAIN_STEPS> \
  --acp.enable=true \
  --acp.indicator_dropout_prob=0.3 \
  --output_dir=outputs/train/<RUN_NAME>

Model & Community

Links and Contact

Release Status

  • Hugging Face model: coming soon
  • Hugging Face dataset: coming soon
  • Canonical version tags will be pinned once published

Citation

BibTeX

@misc{evorl2026,
  title        = {Evo-RL: Continuous Open-Source Real-World RL on SO101 and Beyond},
  author       = {Evo-RL Contributors},
  year         = {2026},
  howpublished = {\url{https://github.com/MINT-SJTU/Evo-RL}}
}