Demo Day 是 2026-09-13，还有 6 天。这一次和前三次都不同：我们全员不在现场，机器人由主办方搬到场地。下面每一条都是按这个约束推导的，不是按"如果有人在"写的。所有"会变的状态"数字取自 STATUS.md（机器生成，2026-09-06 10:09），不是手写估计。
| 操作者（孟苏阳） | 中国，不到现场，已与主办方确认 |
| Ville | 台湾，远程 |
| Demo Day 当天现场 | 队里预计一个人都没有 |
| 机器人 | 芬兰，由主办方搬到场地 |
| 链路（实测） | Tailscale 直连 P2P，中国↔芬兰 RTT 326 ms |
Team name?
Robomates.
1–10, how ready are you?
6。
给 6 不给更高，因为核心能力没验收；不给更低，因为遥操作、安全层、数据链、远程链路都是可重复演示的真东西。
What has to happen this week to make it a 10?
Biggest blocker and who owns it?
What concept will you present?
一台可以从世界任何地方远程操作的调酒机器人，而它在被操作的同时正在学习自己做。
不是两个卖点，是一个飞轮：
远程遥操作 → 每一次操作都被录成示范数据 → 训练策略 → 机器人自己做
↑ ↓
└──────────────── 做不到的地方，人再接管 ←────────────────┘这条链每一环都有当场可指的产物：录制面板、Hub 上的数据集、训练好的 checkpoint、实机试验成绩表。闭环是通的；成功率还不通 —— 而这恰恰是"人还在环里"的理由，不是它的反例。
We help _ do _ so that _
How will you use your time on the stage?
live 为主的混合。
| 时间 | 内容 |
|---|---|
| 0:00–0:45 | 问题 + 一句话定位 |
| 0:45–2:30 | Live：操作者在中国远程操作芬兰的机器人取杯；大屏同时显示控制面板，这一集正在被录进数据集 |
| 2:30–3:30 | Live · 说服力最高的一段：请一位观众上台，随便把杯子挪到任意位置。操作者在 7,000 公里外，不可能事先知道杯子在哪 |
| 3:30–4:30 | 刚才那两段已经变成训练数据 —— 面板上新集入库 → 讲飞轮 → 亮模型和 Hub |
| 4:30–5:45 | 自主尝试（彩蛋）：跑训练好的策略自己抓一次 |
| 5:45–7:00 | 成了是高光；不成就讲真实数据（6/83、判据 8/10、怎么按配置分组统计）+ 谁会买 + 下一步 |
彩蛋放最后是有意的：前 4 分钟已经把分拿到手，自主那一下成不成都不动摇基本盘。"观众挪杯子"那一段不需要自主策略、不需要手眼标定、不需要液体 —— 只需要视频回传和控制链路是活的，而观众心里"这是不是提前录好的"会被当场消灭。
Will 7 minutes be enough?
希望要到 9–10 分钟。我们的形式有跨国切换（现场画面 ↔ 中国操作者），切换本身要时间，而"观众上台挪杯子"的价值就在于不能催。听说有队伍退赛 —— 如果确实有余量，多给的每一分钟我们都会用在 live 上，不用在幻灯片上。
When you arrive how long is setup?
预计 30 分钟，但要说明它怎么拆 —— 因为我们没有人在现场：
| 谁 | 做什么 | 时长 |
|---|---|---|
| 现场协助人员 | 开箱、放上桌子、接电（见下方铁律）、插网线、开机 | ~15 分钟 |
| 远程（孟苏阳） | 之后全部：舵机自检、相机自检、桌高重测、链路自检、试跑一次 | ~15 分钟 |
一旦通电且联网，其余我们全部可以远程完成 —— 控制台是网页的。现场的人真正需要做的只有那 15 分钟的物理接触。
How many people will attend?
现场 0 人。远程 2 人（孟苏阳 · 中国，Ville · 台湾）。
What would you need from us?
Teleop knife fight — 用哪种控制？
我们参加，但必须说清楚：我们只能从中国打，RTT 326 ms。对手如果在现场操作，我们在延迟上先天吃亏 —— 我们接受这个劣势，因为这件事本身就有意思：一场跨 7,000 公里的对抗。控制方式用 VR 或键盘，取决于现场能给的回传画面。如果主办方希望比赛公平，也接受由现场协助人员用主从臂代打。
Anything else, any problems?
base_watchdog.py），能在遥操作进程崩溃或卡死时停下底盘轮子（轮子是速度模式，会一直转下去 —— 2026-08-13 真的失控过一次）。手臂是位置模式，进程死了会停在最后目标位，不是暴露点。关于付费客户反馈（主办方提到有加分） ：目标客户排序是 活动与餐饮服务方 → 酒吧/酒店大堂 → 机器人教育与研究实验室 →（更长期）辅助生活与无障碍。远程操作这一层让第一类客户的账更好算：一名熟练操作者不必到场，可以覆盖多个场地。Demo Day 前会尝试拿到具体反馈。
视频备份：会录，录的是自主抓取成功的那一次。遥操作段不录 —— 它稳。
TEAM: Robomates
1. TEAM & STATUS
Readiness: 6/10.
Solid: full teleoperation (VR + leader-arm), safety layer (self-test PASS: 6/6
synthetic attacks blocked, 0.021% false-kill over 19,453 real frames), the whole
data loop record -> HF Hub -> off-board GPU finetune -> on-robot trial -> results
board, and a live China<->Finland link (Tailscale direct P2P, 326 ms RTT measured).
Not solid: autonomous grasping is NOT accepted. Current-gripper model is 0/6 on
hardware; across all configs 6/83 (7.2%, 95% CI 2.7-15.1%); our 8/10 bar is not met.
To reach 10 this week: (1) an on-site assistant assigned by you, (2) ship our
training table/surface with the robot, (3) add a network-loss failsafe, (4) run the
new left-arm model on hardware, (5) record a backup video.
BIGGEST BLOCKER: nobody from our team will be physically present. Owner: Suyang
Meng (remote operator), who is asking you for an on-site assistant. This now
outranks the grasp success rate: teleoperation can carry the demo, but with nobody
on site the robot cannot even be powered on.
2. ON STAGE
Concept: a bartender robot one operator can run remotely from anywhere in the
world -- and which learns to do the job itself from being operated. Every teleop
session is recorded as demonstration data, trained into the policy, and the human
takes over only where the robot still fails. The loop is closed end-to-end; the
success rate is what we are still working on.
One sentence: We help event and hospitality operators serve drinks with a robot
that one remote operator can run from anywhere in the world, so that every session
it is operated also teaches it to do more of the job alone.
Stage time: LIVE-heavy mix.
0:00-0:45 problem + positioning
0:45-2:30 LIVE: operator in China runs the robot in Finland; the panel on screen
shows the episode being recorded into the dataset as it happens
2:30-3:30 LIVE: an audience member puts the cup anywhere they like. The operator
is 7,000 km away and cannot have known where it would be.
3:30-4:30 those two segments are now training data -> the flywheel -> models+Hub
4:30-5:45 BONUS: the trained policy attempts the grasp autonomously
5:45-7:00 if it works, that is the highlight; if not, the real numbers and how we
measure them + who pays + what is next
No liquid this time: HOLD/POUR are not built, and 6 days is not enough to build a
dispenser that does not spill on stage. We use an empty cup. In our own ranking,
"move the cup and it still finds it" scores higher than pouring anyway.
7 minutes: we would like 9-10 if you have it. Our format switches between the room
and China, and the audience-moves-the-cup segment must not be rushed. Any extra
minute goes into live, not slides.
3. LOGISTICS
Setup: ~30 min. ~15 min of PHYSICAL work by your assistant (unbox, place on table,
power, network cable, switch on) and ~15 min done remotely by us (servo + camera
self-test, re-measure table height, link check, one rehearsal run). Once it is
powered and networked we can do everything else remotely -- the console is a web app.
Attending: 0 people on site. 2 remote (Suyang Meng in China, Ville in Taiwan).
What we need from you:
1) AN ON-SITE ASSISTANT (critical). 15 min of setup, then present during the demo
WITH A HAND ON THE PHYSICAL E-STOP, and to place the cup. We had an arm strike
incident on 2026-08-16; "an operator with a hand on the stop path" is a hard
precondition in our own safety procedure, and with all of us remote only a
person in the room can satisfy it.
2) A WIRED NETWORK PORT or a dedicated router (critical). All control and video
ride this link; a strict-NAT public WiFi would drop our direct connection to a
relay and double the latency.
3) A TABLE -- ideally let us ship our own training table/surface with the robot.
Our vision policy was trained on a fixed table and lighting; changing it will
lower an already low success rate. A mat matching the training surface helps.
4) POWER FROM THE POWER STATION'S 12V CAR SOCKET, NEVER ITS USB-C PORTS. 12V is
not a standard USB-PD voltage; the port cuts out on the first current surge and
it looks exactly like a dead servo. Please pass this to whoever moves/sets up.
5) A picture-in-picture on the big screen showing "Operator: China - live".
6) One cup (we will specify the size). No liquid.
4. OTHER
KNIFE FIGHT: we are in, but we can only fight from China at 326 ms RTT. We accept
the handicap -- a 7,000 km duel is a good story. VR or keyboard, depending on the
video feed you can give us. If you would rather it be fair, your on-site assistant
is welcome to fight for us on the leader arm.
ONE SAFETY GAP WE FOUND OURSELVES AND ARE FIXING THIS WEEK: our watchdog stops the
base wheels if the teleop process dies or hangs (the wheels are in velocity mode and
really did keep rolling once, on 2026-08-13). But if the NETWORK drops while the
process is still alive, nothing fires -- and that is exactly the failure mode for a
fully remote team, with the remote E-stop riding the same dead link. We are adding a
command-timeout that zeroes wheel velocity. Until it lands, the physical E-stop in
the room is the only backstop. That is the technical reason we are asking for a
person, not a formality.
TRANSPORT RISK: our whole approach assumes a fixed scene (same table, same lighting,
5 marked cup positions, table height calibrated to the millimetre). Moving the robot
invalidates all of that at once, so we must re-measure table height on arrival.
PLAN A / PLAN B: the live remote operation is the plan. If the network fails on the
day we switch immediately to a pre-recorded video, which you will have before Demo
Day. The hedge is against the network, not against the concept - what you need to
set up on site is identical either way.
Video backup before demo day: yes, we will record the autonomous grasp succeeding.STATUS.md（机器生成 2026-09-06 10:09）· 仓库内同文 00-admin/CHECKUP-FINAL-2026-09-07.md