code_weights_partial_reproduction_stack · 已核验官方资源
审核码：verified_official; 官方关系：verified_official_code_runtime_and_transformer_weights; 最近核验：2026-07-21代码：Python preprocessing/training/open-loop inference and an Apache-2.0 C++/CUDA runtime are public at fixed commits.
权重：An ungated 6B Transformer is public in three safetensors shards totaling 24,086,435,440 bytes; Wan2.2 VAE and scheduler are external.
数据：The experiment data, 930-episode AutoResearch corpus, task normalization and robot rollout data are not released.
许可证：mixed_apache_and_unlicensed_root；wam.cpp is Apache-2.0 and the Hugging Face model card declares Apache-2.0 for that repository; the Python repository has no root LICENSE, while external Wan2.2 components and user robot data retain separate terms.
依赖：Python 3.11、PyTorch 2.7.1、CUDA 12.6、LeRobot 0.4.4、Wan2.2 VAE/scheduler、C++17/CMake/CUDA and GGUF for wam.cpp
硬件/传感器：NVIDIA A100 author-measured 140 ms C++ path、NVIDIA RTX 4090 author-measured 85 ms C++ path、AgileX PiPER author evaluation、robot-side closed-loop and safety client not released
实机证据：On AgileX PiPER, the author team reports graded averages 0.85 and 0.89 for two six-instruction sets and binary 0.80 for each of two long-horizon tasks, ten trials per instruction or task.
独立复现：No independent fixed-revision latency benchmark, model conversion, physical PiPER deployment or comparable task reproduction was verified.
限制/反面证据：The Python repository has no root license and component licenses do not cover the complete repository.；The public weights contain only the Transformer and still require external Wan2.2 components and task-specific normalization.；The experiment data, robot-side closed-loop client, calibration and safety stack are not released.；All reported latency and physical task results are from the author team; long-horizon evidence covers two tasks with ten trials each.
当前建议：高潜待持续验证；Run a fixed-weight Python-versus-wam.cpp action and latency check before physical shadow mode; do not treat the released Transformer and runtime as an end-to-end reproducible robot policy.
重点标签：主流技术2.0、可能成为新主流、artifact:code-and-weights、latency:85ms-author-result、data-robot-loop:not-released
核验来源：来源 1 · 来源 2 · 来源 3 · 来源 4 · 来源 5
official_runtime_source_update · verified_official_runtime_update
审核码：verified_official_runtime_update; 官方关系：verified_official_apache_runtime_v06_fixed_commit; 最近核验：2026-08-15代码：已核验：Apache-2.0 的固定 0.6 源码可访问，含 C++17 运行时、Python SDK、C ABI、GGUF 转换/检查和模拟器适配层。
权重：未核验到该固定提交可下载的 GWP05/FastWAM source weights 或 GGUF；README 明确将相应制品标为 external 或 soon。
数据：未发布论文实验数据、真实机器人轨迹、标定或机器人侧闭环输入。
许可证：Apache-2.0；Apache-2.0 covers this runtime repository. llama.cpp, Wan/GWP05/FastWAM source models, simulators, datasets and any robot integration retain their own terms.
依赖：CMake 3.22+, C++17 compiler, Python 3.9+ and patch、CUDA toolkit and cuDNN for prediction; CPU path only supports inspection, contracts and development tests、protoc for serving; pinned llama.cpp b9866 is fetched during CMake configuration、external model checkpoint, VAE, tokenizer, normalization statistics and simulator-specific controller inputs
硬件/传感器：NVIDIA CUDA GPU for GWP05/FastWAM prediction、Public source does not include a hardware robot adapter, calibration or safety controller、No independent latency or physical deployment was reproduced in this review
实机证据：No new physical-robot evidence is released by this runtime update; original PiPER results remain author-paper evidence.
独立复现：None verified. The review inspected the fixed source tree and documented build/test entry points, but did not execute GPU inference or simulator benchmarks because the required policy artifacts are not public.
限制/反面证据：README marks source weights and GGUF artifacts as external or forthcoming.；No experiment dataset, real trajectory, robot-side client, calibration or safety stack is included.；Ready simulator support is a repository claim until an independent fixed-artifact benchmark is reproduced.；The update is a runtime/adapter release, not evidence that GigaWorld-Policy, FastWAM or a physical robot stack is end-to-end reproducible.
当前建议：世界动作模型运行时技术改进版·仿真优先·待权重闭包与独立复现；先用固定 commit 的 CPU contract build 与公开/自有模型 artifact 检查接口；只有取得合法权重和同一 simulator 配置后，才比较 FastWAM/GWP05 的时延、成功率与恢复行为。停止条件：没有权重、规范化统计、VAE/tokenizer 或环境控制器时，不把源码构建成功写成策略复现。
重点标签：framework:wam、framework:world-model、artifact:official-runtime-source、evaluation:simulator-adapters、evolution:technical-improvement、community:awaiting-independent-reproduction
核验来源：来源 1 · 来源 2 · 来源 3