code_weights_data_and_evaluation_protocol · verified_hold
审核码：verified_hold; 官方关系：official_project_page_rechecked_github_coming_soon_with_public_small_sample_evaluation_table; 最近核验：2026-08-13代码：official project page still labels GitHub Coming Soon on 2026-08-13
权重：not released on the official project page
数据：processed real/simulated training mix, simulator construction and dream rollouts are not released as a fixed public dataset
许可证：not_disclosed；No method repository, model card or dataset release is available to carry a reusable license.
依赖：Cosmos-Predict2.5 2B、Isaac Sim、Xense tactile rendering、OmniViTac、task-specific Franka Panda and tactile calibration
硬件/传感器：32 H20 GPUs for about 50 hours as reported by the paper、Franka Panda with Xense tactile sensors for author physical tests
实机证据：Official page reports four contact-rich Franka/Xense tasks with ten real trials per policy per task.；For π0.5 plus tactile, the displayed average is 42.5% with expert-only data, 67.5% after the first ViTacWorld rollout round, and 80.0% after a second rollout round; task rates and all four ten-trial denominators are public on the official page.；For the first-round augmented π0.5 plus tactile policy, the official policy-evaluation table reports real-world mean 67.5% versus ViTacWorld-imagined mean 57.5%, with three imagined rollouts per same initial observation and a majority-vote outcome.
独立复现：No independent physical reproduction, released raw trial logs or public end-to-end implementation was located at recheck.
限制/反面证据：Official GitHub remains Coming Soon; no fixed code, weights, processed data, simulator generator or license.；Four tasks with ten trials per policy per task are a small author-only physical protocol on Franka/Xense, not humanoid, wheeled-arm, wheel-legged or cross-lab evidence.；Second-round dream data is generated by the first-round augmented policy, so 80.0% is iterative author-pipeline evidence rather than an independent generalization split.；The evaluator table differs from real execution by 10 percentage points on its displayed average and does not report confidence intervals, failure taxonomy, calibrated uncertainty, intervention rate, safety events or out-of-distribution embodiments.；Manual inspection remains part of dream-data selection according to the paper.
当前建议：视觉触觉世界模型主流技术2.0·有实机评估表·待开放与独立复现；Track the promised release every four hours. If a fixed release appears, validate whether the public preprocessing, simulator, tactile calibration, dream-selection rule and all trial seeds reproduce both the 42.5→67.5→80.0 progression and the imagined-versus-real evaluator gap before awarding community-validation or deployment status.
重点标签：视觉触觉WAM主流技术2.0、真机小样本、世界模型策略评估、合成轨迹迭代、标准泛化性待验证、代码待开放、社区验证待补
核验来源：来源 1 · 来源 2
project_website_source · verified_hold
审核码：verified_hold; 官方关系：verified_official_static_project_repository_no_method_artifacts; 最近核验：2026-08-15代码：The official ViTacWorld GitHub organization exposes a three-entry GitHub Pages repository (README.md, index.html and static/). Its README calls it a plain project page and says final Code URLs remain placeholders; no implementation files are present.
权重：not released; the official project page still renders the GitHub implementation control as Coming Soon
数据：processed 21K-plus training mix, simulator construction and dream rollouts are not released as a fixed public dataset
许可证：not_disclosed；The public static-site repository has no declared GitHub license. It grants neither method-code nor model/data rights.
依赖：Cosmos-Predict2.5 2B、Isaac Sim、Xense tactile rendering、OmniViTac
硬件/传感器：32 H20 GPUs for about 50 hours as reported by the paper、Franka Panda with Xense sensors for author physical tests
实机证据：Paper/project report four contact-rich physical tasks with ten trials per policy per task; no released raw trial logs or independent reproduction.
独立复现：none located at recheck
限制/反面证据：Official project-page implementation button remains Coming Soon；Public repository is website source only；No method code, weights, processed data, simulator generator or license；32-H20 training requirement is not a reproducible deployment closure；Author physical denominator is small and limited to Franka/Xense tasks；Dream-data filtering may confound augmentation gains
当前建议：视觉触觉世界模型主流技术2.0·待开放；Keep a four-hour resource watch. Promote only when an actual method repository or checkpoint is released with license, simulator/data closure and a no-manual-dream-filter reproduction path.
重点标签：visuotactile-world-model、real:small-denominator、code:coming-soon、artifact:static-site-only、artifact:rechecked
核验来源：来源 1 · 来源 2 · 来源 3