code_weights · 已核验官方资源
审核码：verified_official; 官方关系：verified_official_code_and_nongated_checkpoint_releases_after_placeholder; 最近核验：2026-08-13代码：Official repository commit cdd87b6 contains model package, inference server, example client, post-training toolkit and docs; it explicitly excludes the large-scale pretraining pipeline.
权重：Base revision dafcb053 and UniVTAC post-trained revision bdda938 are public and non-gated on Hugging Face. The repository also links separate delta-action and NeoSim checkpoint variants.
数据：NeoData (>30,000 hours / six embodiments in the paper) remains unreleased. Released tooling accepts user-prepared LeRobot latent data and NeoSim/UniVTAC evaluation environments.
许可证：CC-BY-NC-SA-4.0 base/repository; Apache-2.0 metadata on the checked UniVTAC post-trained checkpoint；Repository and base-model release are CC-BY-NC-SA-4.0. The checked UniVTAC post-trained card declares Apache-2.0, but this does not grant rights to the base release, NeoData, Wan2.2 components, tactile hardware integrations or any other checkpoint variant.
依赖：Python environment with PyTorch/CUDA、flash-attn、Wan2.2 VAE, tokenizer and text encoder included by the base-model bundle or fetched under their own terms、websockets and msgpack for the example closed-loop client、NeoSim in a separate environment for the documented simulator evaluation、robot-specific RGB, tactile, action normalization, calibration and safety interface for physical deployment
硬件/传感器：CUDA GPU; minimum VRAM and end-to-end physical control latency are not disclosed、paper pretraining used 128 H800 GPUs、paper physical evaluation used Flexiv/PiPER with InTac S1 tactile sensing; that hardware integration is not supplied as a turnkey stack
实机证据：The paper and official project report eight real contact-rich tasks with 20 trials each, 46.3% macro success, on Flexiv/PiPER with tactile sensing. This remains author-chain evidence, not an independent reproduction.
独立复现：No independent fixed-commit reproduction, hardware bill of materials, calibration report or physical success-log audit found at this review point.
限制/反面证据：No large-scale pretraining pipeline or NeoData release；Base/repository license is noncommercial；Minimum VRAM, end-to-end physical latency, sensor calibration and real-robot safety envelope are not disclosed；Released post-trained checkpoints target UniVTAC/NeoSim rather than a verified turnkey Flexiv/PiPER deployment；No independent physical reproduction
当前建议：触觉世界动作模型颠覆性分支·可做固定制品与仿真复算，真机仍需独立工程验证；优先复算已发布的 NeoSim/UniVTAC 路线，记录传感器缺失与动作归一化失败；未完成公开硬件复现前，不将作者真机平均值当作可迁移性能。
重点标签：触觉世界动作模型颠覆性分支、code:verified、weights:verified、data:closed、community:unverified
核验来源：来源 1 · 来源 2 · 来源 3 · 来源 4
权重 · 已核验官方资源
审核码：verified_official; 官方关系：verified_three_additional_public_nongated_multitask_checkpoint_revisions; 最近核验：2026-08-13代码：Official source was verified previously at fixed commit cdd87b6a141667123ad2c25f452478afdb71e287.
权重：This supplementary review verifies three additional author-linked, non-gated post-training checkpoint revisions.
数据：No N0-TWAM pretraining corpus, paper physical task trajectory bundle or tactile calibration corpus is released.
许可证：Apache-2.0；Each reviewed post-training card declares Apache-2.0. The N0-TWAM base/source, Wan2.2, UniVTAC, NeoSim and robot/tactile inputs retain their own terms.
依赖：Matching N0-TWAM source and base bundle、CUDA/PyTorch、per-task q01/q99 normalization and prompt/camera/tactile/action keys、NeoSim in a separate environment for simulation checkpoints
硬件/传感器：bf16-capable CUDA GPU; exact minimum VRAM and latency not disclosed、matching RGB/tactile/action hardware interface for any physical adaptation
实机证据：The paper reports 20 trials per task across eight contact-rich physical tasks, but this review found no released physical rollouts or independent reproduction.
独立复现：No qualifying independent fixed-commit reproduction was found.
限制/反面证据：No physical rollout dataset or independent reproduction；No complete hardware bill of materials, latency floor or safety validation；Post-training cards do not remove base-model and upstream license obligations；NeoSim checkpoints are simulation-only evaluation resources
当前建议：视觉—触觉世界动作模型主流技术2.0·多任务权重补全；按动作空间和任务域分开复算：UniVTAC delta 必须采用逐任务归一化；NeoSim 两组权重先跑作者说明的每任务 20 个固定种子。不得把 simulation checkpoint 或作者实机表述写成独立实机泛化。
重点标签：触觉世界模型主流技术2.0、视觉—触觉一体化、多任务权重已发布、工程已核验、近期热门
核验来源：来源 1 · 来源 2 · 来源 3 · 来源 4 · 来源 5