权重 · 已核验官方资源
审核码：verified_official; 官方关系：verified_official_public_nongated_checkpoint_revision; 最近核验：2026-08-13代码：Official repository at fixed commit 6e989facc0c452fd3488d75f60bc36411005558c contains preprocessing, training and LIBERO/LIBERO-Plus evaluation entrypoints, but has no repository LICENSE file.
权重：Two public non-gated author checkpoints are listed at the verified Hugging Face revision: dreamwam_joint.pt and dreamwam_uncond.pt.
数据：The training recipe requires FastWAM-prepared LIBERO plus Wan2.2, RAFT, DINOv2, Depth Anything 3 and separately installed benchmark assets; this audit found no DreamWAM-owned physical rollout release.
许可证：MIT；MIT is declared by this Hugging Face model card only. The code repository lacks a root license and all upstream models, data and assets retain separate terms.
依赖：Python 3.10、CUDA 12.x、PyTorch 2.7.1、Wan2.2 TI2V-5B、RAFT、DINOv2、Depth Anything 3、LIBERO and LIBERO-Plus、MuJoCo 3.3.2
硬件/传感器：Eight GPUs for the documented training launch、CUDA GPU for evaluation、Paper evidence uses a controlled dual-arm tabletop platform; no released physical deployment stack
实机证据：The paper reports four standard tasks plus three visual perturbation classes with 30 trials per setting (210 rollouts per model). This review found no fixed-version robot rollout or independent reproduction bundle.
独立复现：No qualifying independent fixed-commit reproduction was found.
限制/反面证据：No code repository license；No physical rollout data or deployment configurations；Requires several large separately licensed upstream models；No independent reproduction
当前建议：世界动作模型主流技术2.0·多模态未来监督的仿真复算可用；先固定 joint checkpoint、四套 LIBERO 和 10,030-task LIBERO-Plus 协议，分别报告 clean 与扰动分母；停止条件是无法获得匹配的 Wan2.2/缓存或版本漂移。不得将作者双臂表写成已公开实机部署。
重点标签：世界动作模型主流技术2.0、RGB+光流+深度+语义、权重已核验、近期热门
核验来源：来源 1 · 来源 2 · 来源 3