代码 · 已核验官方资源
审核码：verified_official; 官方关系：verified_official_staged_release; 最近核验：2026-07-20代码：Visuo-tactile world-model and tactile-aware pi0.5 VLA training/inference modules, conversion scripts, configs and tests are now public.
权重：No TACO-trained checkpoint was verified; all documented weights are upstream Wan/OpenPI components or locally trained outputs.
数据：No paper tactile dataset was verified; users must provide CSV/HDF5 force/tactile sequences and LeRobot conversions.
许可证：Apache-2.0；Root TACO code is Apache-2.0. Gemma, Wan 2.2, robot-DiT and pi0.5 assets have separate upstream licenses/terms.
依赖：Python package environments for the world-model and OpenPI-derived VLA modules、Wan2.2-TI2V-5B、Wan2.2-5B-Robot、pi0.5-DROID base checkpoint、DeepSpeed for the documented training path、LeRobot/HDF5 conversion
硬件/传感器：CUDA GPU capacity sufficient for Wan-style 5B video world-model training/inference、12-dimensional tactile/force sequence input for the released example schema
实机证据：Official paper reports real contact-rich manipulation and a 44-point absolute success improvement over the base policy; no paper checkpoint/data or independent rerun was verified.
独立复现：No independent tactile-hardware reproduction with commit, sensor specification and quantitative metrics was found.
限制/反面证据：The staged source release omits TACO checkpoints, paper data and the full Recognize-Imagine-Label post-training loop.；Required upstream models carry separate terms and significant compute cost.；No independent tactile-hardware reproduction exists.
当前建议：高潜待持续验证；Promote the code release as a material update, but keep performance and end-to-end reproducibility unconfirmed until checkpoints, tactile data and the complete correction loop ship.
重点标签：recency:critical、vla:tactile-post-training、world-model:contact-aware、artifact:staged-code-release、cost:high-gpu、weights-data:not-released
核验来源：来源 1 · 来源 2 · 来源 3 · 来源 4 · 来源 5 · 来源 6
code_partial_release · 已核验官方资源
审核码：verified_official; 官方关系：verified_official_staged_release; 最近核验：2026-07-21代码：Visuo-tactile Wan2.2 training, 12D force handling, HDF5-to-LeRobot conversion and tactile/advantage/knowledge-insulation pi0.5 training and serving are present.
权重：The repository requires external pi0.5-DROID and Wan2.2 bases; no TACO world-model, progress-action or iteration-2 VLA checkpoint was verified.
数据：Schemas and conversion tools are public, but the 300 demonstrations, failed rollouts, imagined corrections, evaluation initial states and force statistics are not released.
许可证：Apache-2.0_with_external_terms；The repository root is Apache-2.0. Gemma, pi0.5-DROID, Wan2.2 and other external checkpoints retain separate terms, and the unreleased project task data has no disclosed artifact license.
依赖：Linux x86_64、Python 3.11、JAX CUDA 12、uv、LeRobot/HDF5、OpenPI/pi0.5-DROID、Wan2.2-TI2V-5B、Wan2.2-5B-Robot
硬件/传感器：8-GPU FSDP training reported without GPU model or duration、Franka Research 3、two Xense 6D force-torque tactile sensors、Intel RealSense D455、SpaceMouse
实机证据：Six real tasks use 40 independent episodes per method and task: base 0.38, iteration 1 0.66 and iteration 2 0.82 average success.
独立复现：No independent fixed-commit tactile-hardware reproduction, confidence interval, cross-sensor result or released task rollout was verified.
限制/反面证据：The correction loop is offline and does not establish online deployment recovery.；All quantitative evidence comes from the author team on one FR3/Xense stack without confidence intervals.；Paper checkpoints, task data, hardware drivers and the complete Recognize-Imagine-Label loop remain unreleased.
当前建议：高潜待持续验证；Prioritize a small in-house tactile HDF5/LeRobot ablation, but do not claim the reported 0.82 success or full post-training loop is reproducible from the current repository alone.
重点标签：颠覆性分支、可能成为新主流、artifact:staged-code-release、vla:tactile-world-model、weights-data:not-released
核验来源：来源 1 · 来源 2 · 来源 3 · 来源 4