Instructions to use Harrysunshine/so101-smolvla-lora-sim-real-10task with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- LeRobot
How to use Harrysunshine/so101-smolvla-lora-sim-real-10task with LeRobot:
# See https://github.com/huggingface/lerobot?tab=readme-ov-file#installation for more details git clone https://github.com/huggingface/lerobot.git cd lerobot pip install -e .[smolvla]
# Launch finetuning on your dataset python lerobot/scripts/train.py \ --policy.path=Harrysunshine/so101-smolvla-lora-sim-real-10task \ --dataset.repo_id=lerobot/svla_so101_pickplace \ --batch_size=64 \ --steps=20000 \ --output_dir=outputs/train/my_smolvla \ --job_name=my_smolvla_training \ --policy.device=cuda \ --wandb.enable=true
# Run the policy using the record function python -m lerobot.record \ --robot.type=so101_follower \ --robot.port=/dev/ttyACM0 \ # <- Use your port --robot.id=my_blue_follower_arm \ # <- Use your robot id --robot.cameras="{ front: {type: opencv, index_or_path: 8, width: 640, height: 480, fps: 30}}" \ # <- Use your cameras --dataset.single_task="Grasp a lego block and put it in the bin." \ # <- Use the same task description you used in your dataset recording --dataset.repo_id=HF_USER/dataset_name \ # <- This will be the dataset name on HF Hub --dataset.episode_time_s=50 \ --dataset.num_episodes=10 \ --policy.path=Harrysunshine/so101-smolvla-lora-sim-real-10task - PEFT
How to use Harrysunshine/so101-smolvla-lora-sim-real-10task with PEFT:
Task type is invalid.
- Notebooks
- Google Colab
- Kaggle
SO-101 SmolVLA LoRA(仿真 + 真机混训,10 个任务)
在 lerobot/smolvla_base 上做 LoRA 微调,训练集是仿真与真机混成的一份:
仿真三个抓放任务 1498 集 + 公开的 SO-101 真机遥操 9 个任务 2200 集 = 3698 集 / 128 万帧 /
10 个不同任务。两侧共有的 Pick up a cube 与 Pick up a can 指令文本逐字相同,
落在同一个任务号下。
三个仿真场景各 50 局的验收
| 任务 | 成功率 |
|---|---|
Pick up a cube and place in the bin |
94.0%(47/50) |
Pick up a small cube and place in the bin |
84.0%(42/50) |
Pick up a can and place in the bin |
98.0%(49/50) |
同一份数据、同一套评测下,全参微调(403M 可训练参数)是 94.0 / 90.0 / 96.0%。 这份适配器只训 9,851,728 个参数(占全模型 460M 的 2.1%)。
挂在哪些层上
上游 SmolVLA 的默认 target_modules 只挂动作专家的 q/v 投影,换机器人本体时学不动 ——
照默认跑到 8000 步,三个评测点全 0.0%。基座里可挂 q/v 的一共三族,这份权重三族全挂:
| 位置 | q/v 数 |
|---|---|
动作专家 lm_expert |
32 |
文本塔 vlm.model.text_model |
32 |
视觉塔 vlm.model.vision_model |
48 |
r=64;state_proj / action_in_proj / action_out_proj / action_time_mlp_in|out
这五个具身投影走 full_training_modules 全量训练,不做低秩近似。
相关
- Downloads last month
- 28
Model tree for Harrysunshine/so101-smolvla-lora-sim-real-10task
Base model
lerobot/smolvla_base