代码 · hold_missing_license
审核码：hold_missing_license; 官方关系：unverified_for_reproducible_open_use; 最近核验：2026-07-20代码：Public repository page and immutable head were inspected, but the promotion gate was not met.
权重：Not promoted or separately verified in this review.
数据：Not promoted or separately verified in this review.
许可证：not_disclosed；README identifies arXiv 2601.21363 and the exact ICLR 2026 paper, but no root license file or GitHub-detected license was found.
依赖：未披露
硬件/传感器：未披露
实机证据：未披露
独立复现：No qualifying independent reproduction with a commit, disclosed hardware and quantitative metrics was established.
限制/反面证据：README identifies arXiv 2601.21363 and the exact ICLR 2026 paper, but no root license file or GitHub-detected license was found.
当前建议：证据观察；README identifies arXiv 2601.21363 and the exact ICLR 2026 paper, but no root license file or GitHub-detected license was found.
重点标签：promotion:withheld、license:missing-or-anomalous、runtime:not-reproduced
核验来源：来源 1
official_code_and_three_model_files_root_license_missing · hold_code_and_weights_license_missing
审核码：hold_code_and_weights_license_missing; 官方关系：verified_official_repository_code_weights_but_root_license_missing; 最近核验：2026-07-20代码：实质代码已发布。提交 32fab57 的 847 个树条目中包含 381 个 Python 文件，覆盖 MuJoCo Playground SAC 预训练、Brax 评测、世界模型训练、微调、JAX 到 Torch 转换和 T1 MuJoCo 播放。
权重：3 个文件端点已匿名核验：T1 rough-terrain zero-shot policy 27,313,236 字节；T1 sim-finetune policy 25,543,932 字节；T1 physics-informed world model 12,945,831 字节。本批未反序列化或运行。
数据：训练脚本可生成 replay buffer，仓库没有发布论文训练 buffer、真实机器人微调轨迹或独立数据集快照。
许可证：not_disclosed；根目录没有 LICENSE、LICENSE.md、COPYING 或 COPYING.md。mujoco_playground 子目录的 Apache-2.0 文件只覆盖其上游代码，不能推定覆盖 LIFT 新增实现、三个模型文件或机器人资产。
依赖：Ubuntu 22.04 and Python 3.10、JAX 0.4.35 CUDA 12 and PyTorch 2.6.0 CUDA 12.4、MuJoCo Playground, Brax, Optax, TorchRL, Hydra, Optuna and Weights & Biases、BoosterGym for the documented Booster T1 deployment
硬件/传感器：README reports testing on NVIDIA H800 and RTX 4090、project page reports several pretraining tasks solved within one hour on RTX 4090、Booster T1 and Unitree G1 models; physical fine-tuning is demonstrated on Booster T1
实机证据：The author project reports zero-shot Booster T1 deployment on indoor/outdoor surfaces and real-world finetuning from 80 to 590 seconds of collected data.；The repository documents model conversion and points to BoosterGym, but does not include a complete guarded real-robot finetuning controller or hardware safety protocol.
独立复现：GitHub issue #2 reports a detailed independent run on Ubuntu 22.04, Python 3.10.20, RTX A6000, CUDA 12.4: 48M-step pretraining reached reward 96 and Brax zero-shot reward 77, while world-model finetuning hit dtype errors, reward collapse/recovery and a 10k-step stall. The author responded that at least 40k steps are needed and pushed a new commit. This is valuable partial reproduction and failure evidence, not a completed comparable finetuning result.
限制/反面证据：代码和三个模型文件缺少根许可证，不能标为已核验开源或可商用。；真实微调轨迹、replay buffer 和安全控制边界未发布。；社区报告表明检查点恢复时 optimizer dtype 易被错误转换，微调可能发生奖励坍塌或长时间卡住。；依赖栈同时包含 JAX/CUDA、PyTorch/CUDA、MuJoCo Playground 与 Brax，版本和显存成本较高。
当前建议：近期高潜：主流强化学习技术 2.0，开放许可待补；大规模 off-policy 预训练与物理先验世界模型微调的组合符合近期高权重路线，且已有社区部分复现；但许可证、稳定微调和真机安全链缺失，在这些问题关闭前只能列为高潜观察。
重点标签：recency:2026-ICLR、关联:主流强化学习技术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:SAC-plus-world-model、artifact:code-and-models-present-license-missing、community:partial-reproduction-with-failure
核验来源：来源 1 · 来源 2 · 来源 3 · 来源 4 · 来源 5 · 来源 6
