quangnd58/evo1-so101-multitask

Evo-1 SO-101 multitask

Evo-1 trained on hungho77/so101-multitask (SO-101 arm, 3 tasks).

Source

• Base: OpenGVLab/InternVL3-1B vision-language backbone (Evo-1 has no robot-pretrained base checkpoint)
• Data: hungho77/so101-multitask (LeRobot v3.0, used as-is)
• Code: MINT-SJTU/Evo-1, branch evo1-lerobot (LeRobot lerobot_train, --policy.type=evo1); this is the stage-2 checkpoint

Details

Cameras	observation.images.top, observation.images.wrist (448x448 inside the model)
State / action	6-D joint positions (shoulder_pan … gripper), LeRobot .pos units
Chunk size	50

Name the robot cameras top and wrist, as in the recording setup.

Tasks (use these prompts verbatim):

• Pick up the banana and place it in the bot, then close the lid
• Pick blue cube and place on red cube
• Pick all cubes and place into cup

Usage

From the evo1-lerobot branch of Evo-1:

PYTHONPATH=evo1_lerobot python -m lerobot.scripts.lerobot_record \
    --robot.type=so101_follower --robot.port=<port> --robot.id=<id> \
    --robot.cameras="{ top: {type: opencv, index_or_path: <i>, width: 640, height: 480, fps: 30}, wrist: {type: opencv, index_or_path: <j>, width: 640, height: 480, fps: 30}}" \
    --dataset.single_task="Pick blue cube and place on red cube" \
    --policy.path=quangnd58/evo1-so101-multitask