代码 · 已核验官方资源
审核码：verified_official; 官方关系：official_substantive_mit_ge_act_robust_wam_source_release; 最近核验：2026-08-14代码：Verified fixed commit has 130 tree entries: GE-Act Robust-WAM training, clean LIBERO and LIBERO-Plus evaluation, DINO target extraction, LTX action/video configurations, inference and web-inference helpers. The FastWAM and LingBot-VA implementations remain explicitly coming soon.
权重：not released: the official README TODO still says pretrained and Robust-WAM checkpoints will be released on Hugging Face.
数据：not released as a project dataset: source expects user-provided LIBERO lerobot-like data with locally precomputed DINOv3 per-frame targets; it does not publish a fixed data snapshot or data card.
许可证：MIT；MIT applies to this repository's declared source. GE-Act, LTX-Video, DINOv3, LIBERO, LIBERO-Plus, Wan/FastWAM dependencies, checkpoints, datasets and robot assets retain their own terms.
依赖：Python environment compatible with GE-Act/LTX-Video、PyTorch 2.7.1、Torchvision 0.22.1、DeepSpeed 0.15.3、Diffusers 0.32.0、Transformers 4.51.3、xFormers 0.0.31.post1、DINOv3 ViT-B/16 for offline target extraction、standard LIBERO and separately installed LIBERO-Plus source
硬件/传感器：NVIDIA CUDA GPUs; official train recipe is single node 8 GPUs, global batch 128 for 50K steps、storage/compute for LIBERO data and per-frame DINO targets、real-robot deployment guide and released policy checkpoint are absent
实机证据：Paper/project report Franka Research 3 visual OOD tests: three tasks, each standard-light and each purple/cyan-light condition with 25 author trials; GE-Act OOD average is reported 57.3% and Robust-WAM 80.0%, InD 80.0% and 82.7% respectively.
独立复现：No qualifying independent fixed-commit physical reproduction or failure report was found in this correction review.
限制/反面证据：Only the GE-Act backbone is released; FastWAM and LingBot-VA folders are not yet released；No official Robust-WAM checkpoint, fixed data snapshot, data card or real-robot guide；LTX/GE-Act upstream environment and LIBERO/LIBERO-Plus are separate installations with incompatible same-name libero modules；Author Franka OOD results are not independent reproduction and cannot be extrapolated to humanoids, wheel-legged robots or tactile manipulation；Repository does not provide a validated latency, safety or intervention protocol for real hardware
当前建议：WAM视觉OOD技术改进版·GE-Act开源复现入口；先以固定提交复算 clean LIBERO 与 LIBERO-Plus；关键指标是同一 backbone、数据版本和 task split 下的 InD/OOD 成功率、端到端延迟和显存。停止条件是无法重建 DINO target、两套 libero 环境混装、无权重却试图宣称论文复现，或真机没有限速/停止/人工接管链。
重点标签：WAM、视觉鲁棒性、语义前瞻、GE-Act、MIT代码、视觉OOD、作者真机、权重待发布
核验来源：来源 1 · 来源 2 · 来源 3 · 来源 4
code_and_weights · 已核验官方资源
审核码：verified_official; 官方关系：verified_official_geact_evaluation_closure_and_public_checkpoint; 最近核验：2026-08-15代码：The fixed 134-file tree contains the GE-Act implementation plus new setup_env.sh, download_assets.py, eval_libero.sh and eval_libero_plus.sh. The wrappers explicitly separate the conflicting LIBERO and LIBERO-Plus Python packages and document source checkout, environment variables and sharding.
权重：The author-linked Haodong082399/Robust-WAM-GE-Act Hugging Face repository is public and non-gated at revision 6a242f6dd4932d7de88106172aebc6d45d819ecc. Its 4,225,696,622-byte diffusion_pytorch_model.safetensors answered an anonymous 64-byte range request with HTTP 206. FastWAM and LingBot-VA checkpoints remain TODO.
数据：No Robust-WAM trajectory, DINO target or physical rollout dataset is released. Evaluation depends on separately downloaded LIBERO and LIBERO-Plus; training additionally needs converted LIBERO-format data and offline DINOv3 targets.
许可证：MIT；MIT covers repository-owned source. The public checkpoint card declares MIT metadata, but GE-Act, LTX-Video, DINOv3, LIBERO, LIBERO-Plus and any converted data retain their own terms.
依赖：Python 3.10、conda、PyTorch and DeepSpeed、GE-Act/LTX-Video stack、Hugging Face Hub、LIBERO、LIBERO-Plus、MuJoCo EGL/OpenGL、DINOv3 ViT-B/16 only for training
硬件/传感器：One CUDA GPU and about 30 GB disk are author-documented for GE-Act evaluation assets、Eight GPUs are documented for the 50K-step GE-Act training recipe、Franka Research 3 remains the paper platform but no released physical bridge, calibration or safety closure was verified
实机证据：The paper reports author Franka Research 3 OOD trials. The updated repository still lists the real-robot deployment guide as TODO; no controller, calibration, safety procedure, rollout log or independent physical replay was verified.
独立复现：No qualifying independent fixed-version reproduction was located. The README's clean-environment run statement is an author claim, not independent evidence.
限制/反面证据：The root README names a FastWAM branch, but the fixed 134-file tree contains no fastwam-RW directory; do not report FastWAM code as verified.；The checkpoint and wrappers have not been executed by this review.；No full training data, DINO targets, FastWAM/LingBot-VA checkpoints, physical deployment closure or independent reproduction is verified.；Clean LIBERO and LIBERO-Plus are separate environments with conflicting top-level module names; mixed environments are not a valid reproduction.
当前建议：世界动作模型主流技术2.0·GE-Act 公开权重与仿真评测闭包；这是从可读源码到可下载 GE-Act checkpoint 和评测脚本的实质工程升级。最小验证应先在分离的 LIBERO 与 LIBERO-Plus 环境对固定 checkpoint 复算；停止条件是依赖无法锁定、checkpoint 不能加载、或与作者声明的 50-trial/扰动协议不兼容。不得据此外推 FastWAM、真机交付或独立复现。
重点标签：世界动作模型主流技术2.0、语义前瞻、视觉OOD鲁棒性、技术改进版、代码已核验、权重已核验、仿真可验证、待独立复现
核验来源：来源 1 · 来源 2 · 来源 3