official_bimanual_vla_code_with_unlicensed_public_model · verified_official_code_only_weights_held
审核码:verified_official_code_only_weights_held; 官方关系:verified_official_repository_code_weights_license_unresolved; 最近核验:2026-07-20代码:已核验开放。提交 8be02cb 包含真实双臂推理、训练、数据适配、RoboTwin 评测、DINOv2/SigLIP/T5-XXL 接口和 flow-matching 策略。
权重:Hugging Face 仓库可匿名下载主模型 pytorch_model.bin(4,137,869,178 字节)、DINO(1,217,515,128 字节)和 SigLIP(3,511,918,424 字节),但页面没有 model card 或许可证元数据,因此不创建已核验权重种子。
数据:论文使用 EgoDex 338K+ 轨迹、194 任务、829 小时;H-RDT 仓库没有分发 EgoDex,也没有提供可直接复制的完整数据缓冲区。
许可证:MPL-2.0;MPL-2.0 覆盖仓库代码文件。Hugging Face 模型、DINOv2、SigLIP、T5-XXL、EgoDex、RoboTwin、机器人资产与训练数据保留各自条款;没有模型卡时不能把代码许可证推定到 4.1 GB 权重。
依赖:PyTorch 2.4.1, DeepSpeed, Diffusers and Transformers、Flash Attention and CUDA-compatible distributed training stack、DINOv2, SigLIP and T5-XXL encoders、custom embodiment dataset adapter and at least 400 GB configured data buffer
硬件/传感器:2B-parameter model and multi-GPU training; issue discussion reports roughly ten days or more for one million steps on author resources、real platforms: dual ARX5, Aloha-Agilex-2.0/Piper and dual UR5 plus UMI、no humanoid-robot embodiment is demonstrated
实机证据:Author results include Aloha towel folding 52% and cup manipulation 64%, ARX5 few-shot average 41.6%, dual-UR5 average 58%, and RoboTwin hard multi-task 87.2%.;These are bimanual robot-arm results rather than humanoid whole-body deployment; cross-embodiment relevance is through VLA/data transfer.
独立复现:Issues #6 and #7 document stale filenames/imports and the need for custom dataset adapters; the author states each embodiment requires fine-tuning rather than zero-shot use. Issue #9 discusses a community estimate of about 500 hours on eight H20 GPUs, while the author reports about ten days-plus for one million steps and recommends MSE/L2 evaluation. No paper-comparable independent endpoint is supplied.
限制/反面证据:公开权重没有模型卡和许可证,匿名可下载不等于合法开放。;EgoDex 338K+ 轨迹不由该仓库分发,完整预训练数据难以按同口径复现。;2B 参数、Flash Attention、多 GPU 和至少 400 GB 数据缓冲区使低成本训练不现实。;每种本体都需要微调,仓库 issue 暴露过时文件名/导入和自定义数据适配成本;没有人形实机验证。
当前建议:主流 VLA 技术 2.0 候选,高算力双臂路线;大规模人类操作数据与双视觉编码器/flow matching 的结合有明显跨领域潜力,但目前应作为高算力双臂研究路线,而不是人形即用基线;权重许可和跨本体真实复现是关键观察点。
重点标签:recency:2025-VLA、关联:主流VLA技术2.0、关联:近期热门、关联:社区验证、关联:可能成为新主流、cross-domain:human-data-plus-bimanual-VLA、artifact:code-open-weights-license-unresolved、cost:high-compute
核验来源:来源 1 · 来源 2 · 来源 3 · 来源 4 · 来源 5 · 来源 6 · 来源 7 · 来源 8