用慢速潜空间扩散动作块保留复杂轨迹建模，并用24FPS触觉/力校正做接触闭环；作者在双Flexiv Rizon 4三任务协议中给出明确分母、公开代码/数据/权重，但尚未跨本体或独立复现。
- 代码 · 已核验官方资源
审核码：verified_official; 官方关系：verified_substantive_official_repository_fixed_commit; 最近核验：2026-08-13代码：released
权重：released separately
数据：released separately
许可证：MIT；Repository source only; bundled upstream notices, ROS/Flexiv/Quest/sensor software and hosted artifacts retain their own terms.
依赖：Ubuntu 22.04、ROS 2 Humble、Python environment with PyTorch 1.13.1+cu117、TactAR Unity app、matching task calibration and data preprocessing
硬件/传感器：Flexiv Rizon 4 with GRAV gripper for the documented default、Meta Quest 3、RealSense cameras、optional GelSight Mini or MCTac tactile sensors、CUDA workstation; README suggests high-performance CPU for GelSight Mini
实机证据：Author paper uses two Flexiv Rizon 4 arms and GRAV grippers across peeling, wiping and bimanual lifting; each test-time variation has ten trials.
独立复现：No independent paper-comparable real-robot reproduction or failure report verified.
限制/反面证据：TactAR Unity companion has no root license at reviewed commit；No container lock or full hardware-calibration closure；No independent reproduction；Arm-only author evidence; no humanoid, wheeled-arm or legged proof
当前建议：高潜待持续验证；Licensed implementation plus explicit real-arm task configuration makes this a worthwhile controlled reproduction target, while independence and hardware closure remain open.
重点标签：视觉触觉主流技术2.0、慢快闭环、接触操作改进版、近期热门、社区验证待补
核验来源：来源 1 · 来源 2 · 来源 3
- 数据 · 已核验官方资源
审核码：verified_official; 官方关系：verified_public_nongated_fixed_revision; 最近核验：2026-08-13代码：released separately
权重：released separately
数据：public non-gated Hugging Face revision
许可证：MIT；Dataset-card declaration; no full-byte download, consent, privacy or third-party asset audit was performed.
依赖：Hugging Face download client、matching RDP preprocessing/configuration、tactile embedding workflow
硬件/传感器：About 198.2 GB listed local storage before preprocessing、CUDA workstation for training
实机证据：Author paper reports 60 peeling, 80 wiping and 50 bimanual-lifting TactAR demonstrations, plus 30 minutes GelSight Mini and 40 minutes MCTac random-interaction data for tactile embedding.
独立复现：Public manifest is verified but no independent end-to-end download, data-quality audit or physical replay verified.
限制/反面证据：Listed byte total is manifest-derived, not a completed content-hash audit；No verified consent/privacy ledger；Arm-only data；Model card MIT declaration must not be expanded to hardware and external dependencies
当前建议：可复现数据候选；A large public, nongated fixed revision supports controlled reproduction, but all content, schema and consent claims require downstream audit.
重点标签：视觉触觉真机数据、双臂接触操作、AR遥操作、MIT声明、需下载复核
核验来源：来源 1 · 来源 2
- 权重 · 已核验官方资源
审核码：verified_official; 官方关系：verified_public_nongated_fixed_revision; 最近核验：2026-08-13代码：released separately
权重：public non-gated Hugging Face revision
数据：released separately
许可证：MIT；Model-card declaration; use requires matching data, code, hardware and upstream dependency checks.
依赖：Matching RDP source/configuration、PyTorch 1.13.1+cu117 path documented by authors、task-specific calibration and tactile representation
硬件/传感器：About 64.1 GB listed local storage、CUDA workstation、documented Flexiv/GRAV/Quest/camera/tactile setup for physical execution
实机证据：Paper reports author dual-Rizon evaluation; hosted checkpoints have not been independently run on the physical protocol.
独立复现：No independent checkpoint execution or paper-comparable real-robot report verified.
限制/反面证据：Listed bytes not hash-audited by this review；No container or complete physical calibration closure；No independent paper-comparable rollout；MIT model-card declaration does not override robot SDK or dependency terms
当前建议：可复现权重候选；Fixed public checkpoint revision makes a controlled replay possible, subject to exact hardware, calibration and data-pipeline reconstruction.
重点标签：视觉触觉策略权重、慢快闭环、MIT声明、需真机复核
核验来源：来源 1 · 来源 2