3万小时触觉数据补齐具身智能“手感”！新智具身&复旦报告三连发
Merged summary
TL;DR — 新智具身（NeoteAI）与复旦大学发布三份技术报告及开源资源，以超过 3 万小时视觉—触觉数据和统一触觉表征，提升机器人精细操作、失败恢复与世界模型规划能力。
- NeoData 覆盖 6 类机器人平台、450 项任务和超过 30,000 小时视觉—触觉交互，其中 5,000 小时已开源。
- NeoForce 统一建模形变图像、电容矩阵、力向量等异构传感器数据，学习可迁移的触觉表征。
- N0/Neo-VTLA 可预测未来最多 50 步的触觉变化，并结合失败数据与离线强化学习，改善插接、取钥匙及长时序操作。
- 72 亿参数的 N0/Neo-TWAM 联合预测未来视频、触觉和动作，在仿真及 8 项真实机器人任务中优于所引用的世界模型基线。
注： 量子位与机器之心聚焦该具身智能研究；雷峰网摘要讨论欧盟低价进口关税，显然与本项工作无关。
Sources (3)
3万小时触觉数据补齐具身智能“手感”！新智具身&复旦报告三连发
TL;DR - NeoteAI and Fudan University released three N0 technical reports, models, and part of a large vision-touch robotics dataset. The work positions tactile feedback as core infrastructure for more reliable physical manipulation.
- NeoData contains over 30,000 hours across 450 tasks and six robot platforms; 5,000 hours are open-sourced.
- NeoForce learns a unified tactile representation across different sensor types.
- N0-VTLA predicts tactile changes 50 steps ahead and uses failure data plus offline reinforcement learning to improve manipulation.
- N0-TWAM jointly predicts future video, touch, and actions, outperforming cited world-model baselines in simulation and real-robot tests.
欧盟3欧元关税落地，卖家在「熬」与「转」间寻找生路
TL;DR - The EU’s new fixed tariff on low-value imports is sharply increasing costs and reducing European orders for sellers on Temu, SHEIN, and AliExpress. Sellers must now choose among costlier direct shipping, inventory-heavy overseas warehouses, or hybrid fulfillment models.
- The €3 levy applies per distinct tariff classification within a parcel, potentially adding €6–€9 to a €20 shipment.
- Sellers report order declines of roughly 40%–60% as higher prices weaken conversion and platforms reduce advertising.
- Overseas warehouses avoid per-order direct-import duties but introduce VAT compliance, substantial upfront inventory, storage, returns, and liquidation risks.
- The Y2 model limits inventory exposure by shipping from China after purchase, but it does not avoid the fixed tariff.
用3万小时触觉数据补齐具身智能「手感」，新智具身联合复旦大学统一触觉「方言」
TL;DR - NeoteAI and Fudan University released three technical reports and open-source resources that make tactile sensing a core input for robot learning. Their approach unifies heterogeneous sensors and uses predicted touch signals to improve fine manipulation and world-model planning.
- NeoData contains over 30,000 hours of visual-tactile interactions across 450 tasks and six robot platforms; 5,000 hours are open-sourced.
- NeoForce learns a transferable tactile representation across sensor formats such as deformation images, capacitance matrices, and force vectors.
- Neo-VTLA predicts tactile changes up to 50 steps ahead and reportedly improves plug insertion, key removal, and long-horizon manipulation.
- The 7.2B-parameter Neo-TWAM jointly predicts future video, touch, and actions, outperforming cited world-model baselines in simulation and eight real-robot tasks.