Sereact Cortex:仓储双臂机器人闭环训练,失误按天回到产线验证
开发者记述 Sereact 在 GTC Berlin 展示 Cortex 视觉-语言-动作模型的闭环训练:Zalando 退货分拣实线数据经 Nebius 与 Isaac Lab/Cosmos 扩增后回训,通过者再回站实测;同周 Teradyne 宣布其为官方伙伴。吞吐与成功率尚未公开。
仓储退货分拣要的是「新品类也能抓」,而不是为每件 SKU 单独示教。
据 @_bbelousov 记述:斯图加特 Sereact 在 GTC Berlin 展示其视觉-语言-动作模型 Cortex 的闭环训练管线,合作方包括 NVIDIA、Nebius、CEVA Logistics 与 Zalando。Cortex 已在德国 Greven 与波兰 Świebodzin 的 CEVA 站点驱动双臂机器人分拣 Zalando 退货时装,无需按货品单独训练。
流程上,产线回合(视频 + 运动学)写入 Nebius 存储;Isaac Lab 扩成大量仿真变体,Cosmos 再按服装状态、包装、材质与光照重渲染;Cortex 在结果上后训练并在留出变体上打分——通过则回站实测,失败则针对失误再扩样本。作者转述的主张是:周一的失误可在数天内变成经过测试的更新,而不是数周。管线由编码智能体驱动,工程师审批超过成本上限的运行。吞吐与成功率尚未公开。
同周 Teradyne 宣布 Sereact 为官方伙伴:已有数百台 Universal Robots 手臂运行在其欧洲客户的双臂系统中。公司由 @ralfgulde 与 @TuscherMarc 创立。
对供应链与仓储自动化读者,这是「实线数据 → 仿真扩增 → 回站验证」的工业 VLA 叙事;采购与试点仍需向厂商索取可复核的节拍、破损率与人机协作安全数据,帖文本身未提供这些指标。
Sereact Cortex closed training loop for warehouse robots
How quickly can warehouse robots learn from the mistakes they make on a live line? Stuttgart-based @SereactAI showed its answer at GTC Berlin: a closed training loop for Cortex, its vision-language-action model, built with NVIDIA, Nebius, CEVA Logistics and Zalando.
Cortex already runs dual-arm robots that sort returned fashion items for Zalando at CEVA sites in Greven, Germany, and Świebodzin, Poland, with no article-specific training.
Episodes from that line (video plus kinematics) go to Nebius storage. Isaac Lab turns them into thousands of simulated variants, and Cosmos re-renders those across garment states, packaging, materials and lighting. Cortex is post-trained on the result and scored on held-out variants. A policy that passes goes back to the stations for a real test; one that fails gets targeted variants for its misses.
The claim: a mistake made on Monday comes back as a tested update within days instead of weeks. A coding agent runs the pipeline, and engineers approve any run above a cost limit. No throughput or success rates are public yet.
The same week, Teradyne named Sereact an official partner: hundreds of Universal Robots arms already work in Sereact's dual-arm systems at customer sites across Europe. Sereact was founded by @ralfgulde and @TuscherMarc.