You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
结论:RLT 与 CAP 不是“哪个更精细”的替代关系。RLT 解决的是已有 VLA 已经到达目标附近后,怎样用短时真机在线强化把最后接触阶段做得更快、更稳;CAP 解决的是怎样用三维接触点明确“应该在哪里发生交互”,并把 pick / open / close 做成可跨场景复用的原子策略。对 DexLab,最值得先借的不是立即复现某个模型,而是把精细任务拆成目标锚定、关键接触阶段、交互预算和完整任务四个可独立验收的层次。
用 Ethernet 插接说明两者差别
设想机器人已经抓住网线水晶头,并把它移动到接口附近:
如果失败是“找错接口、接触目标不明确、场景杂物改变后对不准目标”,更接近 CAP 处理的接触目标条件化问题。
RLT and CAP solve different bottlenecks. RLT locally improves a VLA's precision-critical phase through lightweight online actor–critic learning over an RL token and VLA reference action chunks. CAP conditions atomic utility policies on a 3D contact anchor to remove ambiguity about where interaction should occur. The reported 400–1,000 RLT episodes are online episodes, not demonstrations; their cost must be interpreted together with phase duration, robot-motion time, resets, interventions, wall-clock time, failures and wear. CAP uses 20,365 demonstrations over Pick/Open/Close and does not provide Ethernet or 1 mm insertion evidence.
DexLab should first separate target anchoring, the critical contact phase, the interaction budget and full-task evaluation. A minimal insertion protocol should freeze geometry/clearance, phase boundaries, sensing and compliance; report success, time, retries and forces; compare critical-phase and full-task evaluation; and audit scorer counterexamples. RLT-style refinement is a candidate only when the base policy already reaches the target and fails in the final contact phase. CAP-style anchoring is a candidate when the main failure is spatial interaction ambiguity. Combining them remains an unverified proposal.
核查范围与证据边界
核查日期:2026-11-12。单一 AI 助手整理并自审,没有独立多方评审。
RLT:arXiv 2604.23073、官方论文 PDF 与项目页;论文数字均为作者报告,本轮没有运行策略、真机或社区复现。
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
Uh oh!
There was an error while loading. Please reload this page.
结论:RLT 与 CAP 不是“哪个更精细”的替代关系。RLT 解决的是已有 VLA 已经到达目标附近后,怎样用短时真机在线强化把最后接触阶段做得更快、更稳;CAP 解决的是怎样用三维接触点明确“应该在哪里发生交互”,并把 pick / open / close 做成可跨场景复用的原子策略。对 DexLab,最值得先借的不是立即复现某个模型,而是把精细任务拆成目标锚定、关键接触阶段、交互预算和完整任务四个可独立验收的层次。
用 Ethernet 插接说明两者差别
设想机器人已经抓住网线水晶头,并把它移动到接口附近:
这三个失败层可以相互叠加,但不应合并成一个笼统的“1 mm 精度”。公差、位姿误差、接触后的搜索范围、控制分辨率和任务成功窗口是不同量。RLT 论文把四项任务概括为毫米或亚毫米精细操作,并明确说明螺钉任务需要亚毫米对准;它没有为 Ethernet 任务报告一个独立的“1 mm 公差”测量。
RLT 的 400–1000 episodes 应怎样理解
RL Token(RLT)先用每项任务 1–10 小时遥操作示范适配基础 VLA 并训练 RL token,然后冻结 VLA,用一个小型 off-policy actor–critic 在 RL token、机器人本体状态和 VLA 参考动作上学习。actor 不从整个动作空间盲搜,而是在参考 action chunk 周围改进;BC 正则、reference-action pass-through 与 action chunking 都是论文消融中重要的稳定因素。论文方法与 Algorithm 1
论文报告的 400–1000 是按任务难度变化的在线训练 episode 数,不是 400–1000 条预先收集的示范。每个完整任务约 30–120 秒,而强化学习重点处理的 critical phase 约 5–20 秒;作者报告各实验的实际机器人运动数据约 15 分钟至 5 小时,另有 reset、人工标注、干预与系统开销。Ethernet 的公开曲线在约 15 分钟 critical-phase 数据后达到约 400 次成功/10 分钟的吞吐;项目页同时说明整次训练约两小时。实验设置 · 官方项目页
因此,不能只用 episode 数判断“在线 RL 数据太多或太少”。至少要同时记录:
RLT 在 Ethernet 上的基础 VLA 本来就接近高成功率,主要收益是减少犹豫和重试、提高 critical phase 吞吐;它不是“用数百次失败才从零学会插接”。论文也明确保留了人工终局奖励、干预纠正,以及在基础策略和 RL 策略之间切换的依赖。
CAP 能借什么,不能替 RLT 证明什么
Contact-Anchored Policies(CAP)把语言目标替换为三维 contact anchor。训练时从实际接触帧回溯标注接触点;部署时由人工或 VLM 在图像中给出目标像素,经深度反投影为三维点,再用机器人正向运动学持续跟踪。论文的 contact-anchor 消融在 Close 任务上从 58% 提升到 96%,说明明确交互位置确实能减少策略歧义。
但 CAP 不是低数据在线 RL。公开数据包含 20,365 条示范、约 23.1 小时,覆盖 Pick、Open、Close 三种原子技能;论文和公开 EgoGym checkpoint 也只覆盖这三类任务。CAP 没有报告 Ethernet、peg-in-hole 或 1 mm 插接实验,也没有给出接触锚点等同于机械装配公差的证据。把它称作“更高精度的插接方案”超出了当前公开验证范围。
CAP 对 DexLab 真正有价值的还有另一点:EgoGym 没有被当作“仿真成功就能上真机”的证明,而是用于多 checkpoint 筛选和失败类型分析。论文用四个 Pick checkpoint 做 5,000 次仿真 episode 与 250 次真机运行的单盲对照,再根据 low-lift 等失败修改数据处理。这支持“仿真帮助发现策略退化”的开发环,但该相关性只覆盖论文测过的 Pick 条件,不能外推到精密插接。
对 DexLab 的取舍
先借:精细任务的四层验收
再做:一个不依赖新模型的最小插接协议
DexLab 可以先把一个 connector / peg 插接任务写成评测协议,而不是立刻接入 RLT 或 CAP:
只有当基础策略能稳定到达目标附近、主要失败确实位于最后接触阶段时,RLT 式局部在线强化才是合理候选。只有当失败主要来自“在哪里交互”的歧义时,CAP 式接触锚定才值得优先。两者也可以组合成“contact anchor 定义目标、RL token actor 改进局部动作”,但这是尚未被两篇论文共同验证的研究提案。
当前判断
不建议把 CAP 当作 RLT 的低样本或更高精度替代。 对当前 DexLab,优先沉淀精细操作的阶段语义与完整成本账本:先回答失败发生在目标锚定、接触后动作、物理模型还是评分,再决定需要接触锚点、在线强化、力/触觉闭环或更可信的仿真。
什么会改变这个判断:
English summary
RLT and CAP solve different bottlenecks. RLT locally improves a VLA's precision-critical phase through lightweight online actor–critic learning over an RL token and VLA reference action chunks. CAP conditions atomic utility policies on a 3D contact anchor to remove ambiguity about where interaction should occur. The reported 400–1,000 RLT episodes are online episodes, not demonstrations; their cost must be interpreted together with phase duration, robot-motion time, resets, interventions, wall-clock time, failures and wear. CAP uses 20,365 demonstrations over Pick/Open/Close and does not provide Ethernet or 1 mm insertion evidence.
DexLab should first separate target anchoring, the critical contact phase, the interaction budget and full-task evaluation. A minimal insertion protocol should freeze geometry/clearance, phase boundaries, sensing and compliance; report success, time, retries and forces; compare critical-phase and full-task evaluation; and audit scorer counterexamples. RLT-style refinement is a candidate only when the base policy already reaches the target and fails in the final contact phase. CAP-style anchoring is a candidate when the main failure is spatial interaction ambiguity. Combining them remains an unverified proposal.
核查范围与证据边界
2604.23073、官方论文 PDF 与项目页;论文数字均为作者报告,本轮没有运行策略、真机或社区复现。2602.09017、官方项目仓库与 EgoGym 公开入口;本轮为论文与源码入口静态核查,没有运行 checkpoint。All reactions