这次我想把Post Training里那几秒人工接管再往下拆一层。 @axisrobotics 公开的实验任务，是让机械臂把一块豆腐放到盘子里，Policy先自己执行，出现失败趋势时由人短暂接手，修正后再把控制权交还给Policy。 这里的Correction不是完整示范，而是截取Policy出错位置附近的恢复动作。 ▣ 实验最初收集了660条Correction，第一关是Replay Check，整段操作需要在模拟环境里重新执行并完成任务，660条里有496条通过。 后面还要确认人工接管并非多余，经过Redundancy Filter后剩下416条，其中414条满足短片段测试条件。 到了Recovery Test，人工只修那一小段，控制权重新交回Policy，机器人还要能继续把任务做完。 真正留下的只有161条，平均每段约0.8秒。 ◇ 这套筛选也改变了我的操作习惯。 有一次夹爪已经蹭到物体边缘，我接手后只往侧面调整了一点，把夹爪重新对准目标，确认位置没问题后就点Return to Policy，让机器人自己完成后面的抓取和移动。 ▹ 161条Correction进入训练后，Axis用同一批160个起始状态测试新旧Policy，原Policy完成66次，三次训练分别完成78次，80次和84次。 官方结果能确认完成次数增加，但没有单独拆出具体减少了哪一类失败，我更愿意把它看成Policy整体任务成功表现得到改善。 另一组480个新起始状态里，原Policy完成192次，直接学习完整人类轨迹后降到176次。 换成经过验证的短Correction，Axis还比较了40条，80条和161条三个规模，随着有效修正增加，整体表现继续改善，161条版本三次分别完成236次，235次和225次。 ◈ 现在再看这套机制，我发现Axis筛选Correction时，看的是这段修正到底解决了什么问题，而不是单纯看它持续了多久。 先Replay确认任务能够完成，再判断接管是否发生在Policy确实无法处理的位置，还要经过Recovery Test，验证修正后机器人能否自己继续完成任务。三层筛选都通过，这段Correction才会被留下。 模拟环境里的效果已经有数据支撑，下一篇我想把视线转到真实机械臂上，看看这些经过筛选的修正离开浏览器以后还能不能继续发挥作用。 实验数据来源

$VICR 受到利好消息刺激，盘后股价涨幅达到11% VICOR 官宣一家领先的AI OEM厂商获得其VPD专利产品的授权。这对VICOR而言，是重大突破。（尤其考虑到VICOR目前正挥舞专利大棒，向诸多竞争对手发起了新一轮诉讼） ============================= Vicor licenses VPD to a leading AI OEM Sep 16, 2026 at 4:00 PM EDT ANDOVER, Mass., Sept. 16, 2026 (GLOBE NEWSWIRE) -- Vicor has granted a non-exclusive license to a new OEM licensee with the right to procure Vertical Power Delivery (VPD) modules covered by Vicor patents from unlicensed suppliers. VPD enables high-performance AI compute and network processors fueled at a high current density from a low voltage source. VPD systems, consisting of a “multi-cell converter” and a VPD network, enable breakthrough performance by avoiding the “the last inch problem” of Lateral Power Delivery (LPD). VPD systems are protected by foundational patents to innovations conceived during a decade of pioneering research conducted by Vicor in the U.S. VPD licensees can source VPD solutions from unlicensed global suppliers, whose multi-cell Voltage Regulators (VR) or Integrated Voltage Regulators (IVRs) are stacked in a multi-layer system. VPD solutions using high current density and high gain MCM™ current multipliers manufactured as ChiPs™ (Converters housed in Package™) in a Vicor fab are also available to licensees. MCMs enable substantially higher total system efficiency, are much thinner and mechanically and thermally adept. By procuring VPD modules for high-performance AI accelerators from Vicor, licensees may earn substantial discounts on licensing royalties for commodity VPD modules from unlicensed suppliers. The Vicor licensing practice provides OEMs and hyperscalers with multi-source flexibility, supply chain redundancy, and scalability, including from U.S. based Vicor fabs. As high-performance compute and network processors migrate to VPD, unlicensed OEMs and hyperscalers may need a license to secure their supply chain. Potential licensees may expect a structured licensing practice providing much lower royalty rates for early adopters. Supply agreements with infringing suppliers may contain indemnification provisions for recovery of royalties and monetary damages, but cannot overcome importation bans.