Modelpublic

MikeWu597/URIS_TurboVLA_PiPER

sourceHugging Faceupdated 2mo agoView on Hugging Face

0likes11downloads

Model Card

URISTurboVLAPiPER

TurboVLA在PiPER真机遥操数据上微调 10000 步的权重。基座为 MikeWu597/URIS_TurboVLA_Pretrained，代码采用 H-EmbodVis/TurboVLA。

文件

文件	说明
`model.safetensors`	FP32主权重，106.6M 参数（DINOv3 ViT-B 85.7M + 策略头 20.9M）
`ema_model.safetensors`	EMA版本的主权重（decay=0.999）
`config.json`	模型结构超参 + 微调配置 + 7 条任务指令
`piper_stats.json`	动作/状态归一化统计
`piper_bert_text_cache.pt`	任务指令缓存
`loss_curve.png`	微调 loss 曲线（10k 步）

微调配置

数据采用piper_orbbec_lerobot_joint（LeRobot v2.0）350 episodes，冻结 DINOv3，batch 32 * accum 2（effective 64），lr 2e-5（warmup 500），10k 步

[image]

加载方式

python

import torch
from safetensors.torch import load_file
from turbovla.models.turbovla import build_turbovla

class A:  # 与 config.json 一致
    LOCAL_DINOV3_PATH = "/path/to/dinov3-vitb16-hf"
    text_hidden_dim = 768; hidden_dim = 256; nheads = 8
    dim_feedforward = 2048; enhancer_inner_dim = 1024; max_text_len = 256
    vla_feature_enhancer_layers = 6; action_dim = 7; chunk_size = 12
    state_dim = 8; num_state_tokens = 2
    text_dropout = 0.0; fusion_dropout = 0.0; fusion_droptath = 0.1
    local_files_only = True; freeze_vision_encoder = True

model = build_turbovla(A())
model.load_state_dict(load_file("model.safetensors"))
model.eval()