代码 · 已核验官方资源
审核码：verified_official; 官方关系：verified_official_framework_source_update_wala_wan_lerobot_v3_and_split_evaluation; 最近核验：2026-08-15代码：MIT fixed commit a27d2afbb315b1378237569e79f1397857ec6a29 is a material update over the initial e2b5ca revision: it adds the Wan2.2-TI2V-5B WALA-Wan policy, Wan training/configuration launchers, LeRobot v3 shard indexing, depth-cache preprocessing, an H20 RoboTwin evaluation launcher and documented split policy-server/simulator operation.
权重：not released. The author README says complete training data, model checkpoints and RoboTwin/RoboCasa benchmark checkpoints will be released only after the new 25,000-hour pretraining run completes.
数据：not released. The repository supports user-supplied LeRobot v2.1/v3.0 data but excludes paper robot demonstrations, action-free human videos, simulator assets and training outputs.
许可证：MIT；MIT covers repository-owned WALA source. Qwen3-VL, Wan2.2, DINOv3, Depth Anything V2, RoboCasa-GR1, RoboTwin, cuRobo and user datasets require their own compatible terms.
依赖：Python 3.10、PyTorch 2.6.0 CUDA 12.4 build、DeepSpeed ZeRO-2、Qwen3-VL-4B-Instruct or Wan2.2-TI2V-5B Diffusers release、DINOv3、Depth Anything V2、LeRobot v2.1 or v3.0 compatible data、separate RoboCasa-GR1 and RoboTwin environments、RoboTwin cuRobo v0.7.8 compatibility when required
硬件/传感器：CUDA workstation for training/inference、paper reports 70 ms inference on RTX 4090、current repository includes a RoboTwin H20 evaluation launcher but does not establish a minimum or paper-equivalent hardware budget、physical robot, paper demonstrations and calibration stack remain absent
实机证据：The paper reports four author real-manipulation tasks with 30 trials per configuration; WALA with 200 robot demonstrations plus 400 action-free human videos reports 100/120 successes, and one bread OOD task reports 9/10 after adding 400 videos.
独立复现：No qualifying independent physical reproduction or fixed-checkpoint evaluation found. The current source supports simulation-oriented rebuilds only after the user supplies third-party checkpoints, datasets and simulator assets.
限制/反面证据：No WALA model checkpoint or 25,000-hour pretraining output；No paper robot demonstrations or action-free human-video corpus；No RoboCasa/RoboTwin assets bundled；Wan2.2 and other third-party model rights are separate；RoboTwin interface requires a specific cuRobo v0.7.8 compatibility path on some revisions；No independent physical reproduction
当前建议：主流世界动作/VLA技术2.0·代码迭代版·仿真可复算，真机待复现；先用公开源码在自有 LeRobot/RoboTwin 或 RoboCasa 数据上复算，记录当前观测、动作归一化、策略服务器延迟和 cuRobo 兼容失败；缺论文权重、数据与独立真机日志时，不把作者 120 次结果升级为可迁移实机结论。
重点标签：WAM、VLA、主流技术2.0、人类第一视角视频、WALA-Wan、LeRobot-v3、RoboTwin、RoboCasa、RTX-4090-latency、代码开放、权重待发布、数据待发布、社区验证待补
核验来源：来源 1 · 来源 2 · 来源 3 · 来源 4