搬场地 · 标定 / 桌高 / 离开前采集 · 对照硬编码 IK 链与 SmolVLA 链 · 2026-09-11

搬到新场地要重做什么：桌面高度、标定清单、离开 Space 21 前必须采的数据

整套系统建立在「场景固定」上：同一张桌子、同一头姿态、同一光照、五个杯位，桌高标到毫米。搬场地时这些前提同时失效。这页把每一个写死在代码或数据里的场地量列出来，分成两条链对照 —— 硬编码 IK 链（安全层、归位、放杯、视觉伺服）改常量就能跟上；SmolVLA 链（数据集、checkpoint）改不了，只能让新场地长得和旧场地一样，或者重录。结论先行：桌面相对肩轴的高度必须复现到 ±1 cm，否则学习策略作废；离开前必须量一次「地面到桌面」的物理高度，现在仓库里没有这个数。

来源：configs/teleop_safety.yaml · configs/table_calibrations.json · configs/head_presets.json · scripts/measure_table.py · teleop_config.py z_floor()/y_floor() · policy_safety.py · cam_to_arm.py · visual_servo.py · teach_grasp.py / TEACH_GRASP.md · cup_locate.py · run_policy_trials.py（CAMERAS、--preview、start_poses_for）· usb_cameras.py · fix_hotspot_internet.sh · ROBOTIC-WORKING-HANDBOOK.yaml（P-03/P-06/V-01/V-06/V-07/S-01）· GRIPPER-SWAP-CHECKLIST.md · CURRENT-STATUS.md · 00-admin/DEMO-DAY-AWARDS-AND-LOGISTICS-2026-09-10.md · ~/.cache/huggingface/lerobot/calibration/

给现场同学：离开前做这几件，其他都不用管

用卷尺量四个数，拍照。地面到桌面；地面到左肩关节轴、右肩关节轴（手臂根部那颗转轴的中心）；地面到头相机镜头；机器人前边缘到桌子边。顺手量一下桌子长宽。把数字写在纸上拍一张，发到群里。
在地上贴胶带标出四个轮子的位置，从上面拍一张。
给三根相机线贴标签并拍照：左手腕相机、右手腕相机、头相机各插在集线器哪个口，标签写清楚，拍一张能看清插口的照片。到了新场地照原样插回去。
搬之前：告诉远程把力矩卸掉，手臂软了以后扶着放平，关电源，头用手轻轻压到最低，相机线不要拔。
到了新场地：先量地面到桌面、地面到肩轴，和第 1 条比。桌面到肩轴要在 13.5 厘米左右，差超过 1 厘米就在机器人脚下垫板子（垫机器人，别动桌子）。然后开电源、连网，剩下的远程来做。

以上大约 40 分钟。下面的内容是给远程操作的人看的，现场不用读。

0 · 一句话版

桌子。所有几何都以肩轴为原点，桌面在肩轴下方 0.1349 m（左 z_min 0.061 − 爪长 0.1959；右 0.065 − 0.1999，2026-08-24 双臂各量、双重验证，两臂差 4 mm）。新场地要么把桌面（或机器人脚下的垫高）调到肩轴到桌面 = 13.5 cm ± 1 cm，要么接受学习策略从头再来。到场第一件事是用 measure_table.py 复测，不是信旧值。

硬编码 IK 链：8 个配置文件、约 2 小时，全部有现成脚本可重测（除 table_datum.json 缺失和底盘比例系数从没标过）。

SmolVLA 链：checkpoint 里写死了旧桌高、旧头姿态（每集内头部标准差 0.0000，对头动零容忍）、旧光照（两批 G/R 1.28 和 0.98）。任何一项变了都是分布外，唯一的修法是重录 ≥50 集再训（4090 上 1～2 天）。所以搬场地的目标不是「重标」，是「把新场地做成旧场地」。

离开前必采（第 4 节，约 2.5 小时）：物理尺寸卷尺量一遍并拍照；参考配置下 10 集预注册基线；头姿态五张参考图 + 空桌底片；--preview 通过日志；所有配置打包并提交；两个数据集和 checkpoint 确认在云上。

1 · 哪些量写死了场地 —— 按物理依赖分组

依赖	文件 · 值	谁在用	怎么重测	耗时
桌面高度（相对肩轴）	`configs/teleop_safety.yaml` arms.left.z[0]=0.061 / right 0.065；gripper_len_m 0.1959 / 0.1999；home_pose z 0.17（开机自检要求指尖离桌 ≥100 mm，桌高变 home 也变）	安全层每帧（policy_safety.py 调 z_floor/y_floor 按爪俯仰角算地板）、归位、遥操作	`measure_table.py --arm left --port …`：卸力矩、手把指尖按在杯子会站的位置、FK 读腕高；两臂各量；面板或 YAML 手填。`--pivot` 顺便反推爪长	30 min
	`configs/table_calibrations.json` 最新 20260824-064847，含头 pan/tilt 和三张照片	历史与回滚（robot_server calib_restore）	新场地追加一条，不改旧的；服务端会拒绝两臂差 >20 mm 或散布 >10 mm 的标定	随上项
	`cam_to_arm.py table_datum()` → configs/table_datum.json 不存在，静默回退 0.20 m	头相机像素→臂坐标	无脚本，要写一个（或先记下再说）	—
	`configs/place_pose.json`（10 集示范，2026-09-07）、`pour_pose.json`、`trial_start_pose*.json`（每集第一帧中位数，arm_home.start_poses_for 可离线重算）	放杯、倒酒、试跑起手位	关节角是绝对量，桌高变了不能平移；place/pour 要重教 10 集；起手位从数据集来，数据集不变它就不变	place 30 min
	`visual_servo.py`「杯子在固定平面上」；高度来自接近位姿	视觉伺服抓取	接近位姿随桌高重教	随 teach_grasp
	过期文案：robot_server.py 还写着爪长 0.159、右 z_min +0.02、桌面 −0.139，与 YAML 矛盾	面板说明文字	重测时顺手改	5 min
头相机姿态	`configs/head_presets.json` cup-grasp-v0 = pan 2003 / tilt 2619（操作者 2026-09-04 定为标准视角）。注意 2619 比舵机 range_max 2617 高 2 tick	试跑锁头、录制、robot_effectors HEAD_PRESET	`head_aim.py --preset cup-grasp-v0` 后看图：/api/reference/head 五张参考图（不同杯位）必须都能叠上（手册 P-03：「相似亮度不等于在看桌子」）	15 min
	`configs/head_reference_{0..4}.jpg`、head_reference_cup_grasp.jpg（2026-08-23）	头对准卡片	这是 Space 21 的照片，搬走即误导；新场地重拍，旧的改名留档	10 min
	cam_to_arm_left/right.json、servo_jacobian_left.json（48 样本，2026-09-07，头在 pan 2000 tilt 2615）、wrist_jacobian_left.json、grasp_jacobian.json（2026-09-10）、grasp_demos.json（8 条，09-10）、grasp_tip_offset.json（n=2，std 大于均值，本来就没用）	视觉伺服、teach_grasp 抓取	cam_to_arm 文档：「只在拟合时的头姿态下有效」。`teach_grasp.py` 全流程 12 min（TEACH_GRASP.md）；`visual_servo.py --measure-jacobian --go` 10 min；`cam_to_arm.py --sweep` 30 min	50 min
光照 / 颜色	`cup_locate.py` CUP_H 12–22、S≥80、V≥75、TABLE_TOP 0.18（2026-09-10 白天在头相机上量）；cup_mask.py 蓝/黄阈值（09-11）；finger_locate.py 胶带阈值	杯子定位、抹杯、指尖定位	新场地重拍 12×12 色块量 HSV；桌面必须仍是白的（S<25）	20 min
	头相机 G/R：白天 0.95～0.98、夜间 1.54～1.58；数据集两批 1.28 / 0.98	SmolVLA 输入分布；手册 V-06 夜间成绩单独标	每次试跑 trial_keyframes 自动记 light.gr/lum；新场地第一次试跑先看这个数	0
	cup_mask --mask-fill plate 需要的空桌底片	颜色选杯	头到 cup-grasp-v0，桌上无物拍一张 424×240	2 min
USB 拓扑	相机 by-path 三处写死（robot_server.py:86、run_policy_trials.py:861、camera_view/identify_cameras）：左腕 2.3.2、右腕 2.3.3、头 1.1:1.4。Orbbec 09-03 已经换过一次口；两只腕相机序列号相同，只靠 hub 口区分	录制、试跑、预览	插错口左右腕会静默互换。到场先跑面板 identify_cameras（转一下腕看哪路画面动），再 `run_policy_trials.py --preview` 看相机路径/分辨率 PASS	10 min
USB 拓扑	串口 /dev/ttyACM0/1 按舵机数识别，不按口	所有总线脚本	不用管	0
地面	base_translate.py CAL_SCALE = 1.00「标过之后填这里」，从没标过；WHEEL_MOUNT_DEG 是机器本体的	底盘平移（演示不用）	开 0.5 m 卷尺量；地毯和瓷砖不一样	10 min
网络	Tailscale 100.107.145.111 随机器走；8770/8790/8781/8443/8797/5900 端口不变	远程面板、上传、VR	不用管	0
网络	fix_hotspot_internet.sh 写死 WAN_IF=enP8p1s0（有线 193.167.0.37）、AP_IF=wlP1p1s0、子网 10.42.0.0/24；iptables 不过重启	Quest 热点上网	新场地是 Wi-Fi/DHCP 的话改 WAN_IF 重跑；`verify_hotspot_client.sh` 验	15 min
机器本体（不随场地变）	舵机标定 ~/.cache/huggingface/lerobot/calibration/robots/xlerobot/xlerobot_arms_only.json（2026-08-16）；运动学 L1 0.1159 / L2 0.1350；max_reach_m 0.2459；WHEEL_MOUNT_DEG；pan_sign	一切	只在换舵机后重做（lerobot-calibrate，提示符打 c）。左肩 shoulder_lift 09-09 曾 69 °C 堵转、今天全天 50 °C+，如果搬场前换了它，整条链从舵机标定开始重来：z_min、爪行程、pan_sign 全部重测（SETTINGS-AUDIT 规则）	换舵机 +1 h

2 · 两条链对照：桌高变了各自怎么办

	硬编码 IK 链（安全层 · arm_home · place/pour · 视觉伺服 · teach_grasp）	SmolVLA 链（sep3-clean 54 集 · left-sd-p0.2-b16-u20000@020000）
桌高在哪	一个数：teleop_safety.yaml z[0]，其余从它派生	13,693 帧的关节角和图像里，无处不在
桌高 +Δ 时会怎样	不重测：安全层地板按旧桌算，策略/伺服把指尖按进桌面 3 cm 才被拦（min_wrist_clearance 0.03）。重测后：地板正确，place/pour 目标要重教	它输出的是旧桌面上的绝对关节角：桌高 +1 cm 就在杯壁高 1 cm 处闭爪（多半还行）；+3 cm 指尖顶桌面被安全层 HELD 掐掉整轮；−3 cm 在杯口上方空抓。没有任何参数能补
头姿态变了	cam_to_arm / jacobian 全部重拟合（有脚本）	零容忍：54 集里头部标准差 0.0000。tilt 不对等于「对着墙」，手册 P-03 有过整轮作废的先例
光照变了	cup_locate/cup_mask 阈值重量（20 min）	夜间 G/R 1.54 时成绩要单独标（V-06）；训练集里没有的色温就是分布外，程度未知
杯位范围	由 teach_grasp 示范覆盖，可补示范	120～175 mm 中段五档（cup-positions.json）；左臂闭爪 pan 中位 −13.5°，超出范围就是外推
修复成本	≈2 h，全部现场可做	重录 ≥50 集 + 上传 + 4090 训 20k 步 ≈ 1～2 天；或者不改它，改场地

3 · 桌面高度怎么调 —— 三种情况

新桌子高度和旧的一样（±1 cm）。到场：measure_table 两臂复测 → 值应落在 0.061 / 0.065 ± 0.010 内 → 追加 table_calibrations 记录 → 头对 cup-grasp-v0 看图 → identify_cameras + --preview → 跑 5 集基线和离开前的 10 集比。1 h，策略不动

高度不一样，但能调（升降桌、垫木板、机器人脚下垫板）。目标是「肩轴到桌面 = 13.5 cm」，不是「桌子多高」。用离开前量好的「地面到肩轴」减去新桌面高度，算出要垫多少；垫完按 A 走。垫机器人比垫桌子好：桌面材质和颜色不变（cup_locate 要白桌面），头相机看到的桌面纹理也不变。1.5 h

高度不一样，也调不了。硬编码链按第 1 节全部重测（2 h），演示只走脚本抓取（grasp_cup / teach_grasp 那条线，09-10 已在真机跑通）。SmolVLA 只能新录：到场第一天用 leader-follower 录 50 集（头必须先到标准视角、按录制指南四项验收）、当场 /api/dataset/publish、4090 训一晚。2 天，且成功率回到未知

容差怎么定的：安全层腕部离桌余量 30 mm，超过它策略会被 HELD 掐掉；示范里闭爪高度在杯壁中段，杯高约 9 cm，±1 cm 仍在杯壁上；两臂标定的合格线本来就是差 <20 mm、散布 <10 mm。所以 ±1 cm 是「策略大概率还能用」，±3 cm 是「肯定不能用」。中间那段没实验过，别赌。

4 · 离开 Space 21 前要做的事 —— 分三类

A 是现场才能做、仓库里没有的；B 是可选补跑；C 是仓库里已经有的，只确认备份。2026-09-11 晚复核后，原来放在 A/B 里的桌高读数、参考图、底片、光照色块、三组基线全部挪到了 C。

A · 仓库里没有，只能在现场做（约 35 分钟）

#	做什么	工具	存哪	耗时
A1	卷尺量：地面→桌面；地面→左肩轴、右肩轴；桌面→肩轴（应 ≈13.5 cm，和 YAML 反推互验）；地面→头相机镜头；机器人前缘→桌边；桌子长宽	卷尺 + 手机照片带比例尺	configs/venue-space21-2026-09-11.json + 照片	20 min
A2	机器人停放点：地上贴胶带标四个轮位，拍一张俯视	胶带、手机	同上	5 min
A4	--preview 通过日志（相机路径 / 分辨率 / state 维序），机器空闲时跑 `run_policy_trials.py --preview` 存到 05-training/trials/preview-space21.log；09-11 晚总线被 teach_present 占着，还没跑	train310 python	05-training/trials	5 min
A3	相机线插在 hub 哪个口：贴标签拍照（左腕 2.3.2、右腕 2.3.3、头 1.1）。设备侧的拓扑已存：05-training/trials/usb-topology-space21-2026-09-11.txt（lsusb -t + by-path）	标签、手机	04-reports 照片	10 min

B · 可选补跑（仓库里已有同配置成绩，补是为了样本量）

#	做什么	已有的	耗时
B1	右臂再 5 集，P2～P4，同今天配方	09-11 6/6（trials-20260911-152009.json）	25 min
B2	左臂再 5 集，杯子画面左半近端	09-09 15/20（28 批）、09-11 凌晨 4/4	20 min
B3	夜间头参考图一套（白天那套已有）	凌晨试跑关键帧可当夜间参考	5 min

C · 仓库里已经有，只确认备份（约 20 分钟）

已有什么	在哪	要做的只是
桌高标定与安全包络	configs/table_calibrations.json（20260824-064847，双臂各量、差 4 mm）；configs/teleop_safety.yaml	一条命令 tar 打包到 ~/incoming 下载走；git 由操作者提交（git status 里 20 多个未跟踪文件）。10 min
头预设、参考图、空桌底片	configs/head_presets.json；head_reference_{0..4}.jpg + head_reference_cup_grasp.jpg（08-23）；head_plate_cup-grasp-v0.jpg（09-11）；今天所有试跑的 init 关键帧
光照色块与夜间数据	cup_locate.py 注释（09-10 实测 HSV）；trial_keyframes 每集记 light.gr（白天 0.95、夜间 1.54）
基线成绩	左臂 09-09 15/20、右臂 09-11 6/6、夜间左臂 09-11 02:xx 4/4；全部在 trials-*.json 与 STATUS.md
相机→臂标定、雅可比、示范	configs/cam_to_arm_*.json、servo_jacobian_left.json、wrist_jacobian_left.json；scripts/grasp_jacobian.json、grasp_demos.json、grasp_tip_offset.json
脚本姿态	configs/place_pose.json、pour_pose.json、trial_start_pose*.json、marker_offset.json、pan_sign_measured.json、pan_range_measured.json
舵机标定	~/.cache/huggingface/lerobot/calibration/（机器本体的，不随场地变）
杯位与试跑记录、USB 拓扑	05-training/trials/cup-positions.json、traces/、keyframes/、usb-topology-space21-2026-09-11.txt
数据集	sep3-clean（54 集）、0826sep8-full（73 集）	查 meta/upload.json 为 ok（R-05）。5 min
模型	05-training/…/checkpoints/020000（左、右各一）	确认 HF Suyang99 下有同名 repo。2 min

合计：A 约 40 分钟是硬指标，没有它到了新场地不知道要垫多少；B 可选；C 约 20 分钟。

5 · 到场后的顺序（照着做，约 2 小时 + 基线）

开机前：卷尺量地面→桌面、地面→肩轴，和第 4 节 #1 比，决定垫不垫（第 3 节）。
check_arms.py：17 颗全应答、温度正常（左肩 <50 °C 再开跑）。
identify_cameras → --preview PASS；对不上就换口重跑，别改代码里的 by-path。
measure_table.py 两臂 → 追加 table_calibrations → 改 YAML z[0]（和 home z 如果桌面动了）。
head_aim --preset cup-grasp-v0 → 五张参考图叠图确认 → 拍新底片。
arm_home 双臂归位自检通过（P-01）。
光照：第一次试跑看 light.gr；>1.3 就按夜间标；cup_locate/cup_mask 阈值按需重量。
teach_grasp.py 12 分钟重教（脚本抓取线）；place_pose 重教（如果演示要放杯）。
基线 5 集，六项配置与第 4 节 #3 完全一致；和离开前的 10 集并排写进 STATUS。
Quest 热点：WAN_IF 改成新场地的口，verify_hotspot_client.sh。

6 · 已知缺口（这页没解决的）

没有运输文档：断电顺序、卸力矩、轮子锁、线缆固定。建议：check_arms --release → 关电 → 头手动放到 tilt 最低 → 相机线不拔（拔了就要重做第 4 节 #7）。
急停对 leader-follower 无效（HANDOFF-2026-09-09-NIGHT）：新场地录数据时这个 bug 还在。
左肩 shoulder_lift 09-09 69 °C 堵转、今天持续 50 °C+ 欠力 6～7°：搬前换不换要定；换了走「机器本体」那行的全套重标。
table_datum.json 缺失、CAL_SCALE 没标过、tilt 2619 超 range_max 2 tick：三个老问题，搬场地是顺手清掉的时机。
本页所有耗时是估算（证据等级：没跑过），脚本存在与参数已核对（证据等级：跑测过 ls/grep）。

数值出处见各行文件名；桌高 −0.1349 m 与 z_min 来自 teleop_safety.yaml 与 table_calibrations.json 20260824-064847；页面写于 2026-09-11 下午，机器在 Space 21。

← reports index

Venue move · calibration / table height / what to capture before leaving · hard-coded IK chain vs SmolVLA chain · 2026-09-11

Moving to a new venue: table height, the recalibration list, and the data to capture before leaving Space 21

The whole stack assumes a fixed scene: the same table, the same head pose, the same lighting, five cup positions, table height calibrated to the millimetre. A venue move breaks all of these at once. This page lists every venue-bound quantity baked into code or data and splits it into two chains — the hard-coded IK chain (safety layer, homing, place, visual servo) follows a constant change; the SmolVLA chain (datasets, checkpoints) cannot be changed, so the only options are to make the new venue look like the old one or to re-record. Verdict first: the table height relative to the shoulder pivot must be reproduced to ±1 cm or the learned policies are void; and before leaving, measure the physical floor-to-table height once — the repository does not contain that number.

Sources: configs/teleop_safety.yaml · configs/table_calibrations.json · configs/head_presets.json · scripts/measure_table.py · teleop_config.py z_floor()/y_floor() · policy_safety.py · cam_to_arm.py · visual_servo.py · teach_grasp.py / TEACH_GRASP.md · cup_locate.py · run_policy_trials.py (CAMERAS, --preview, start_poses_for) · usb_cameras.py · fix_hotspot_internet.sh · ROBOTIC-WORKING-HANDBOOK.yaml (P-03/P-06/V-01/V-06/V-07/S-01) · GRIPPER-SWAP-CHECKLIST.md · CURRENT-STATUS.md · 00-admin/DEMO-DAY-AWARDS-AND-LOGISTICS-2026-09-10.md · ~/.cache/huggingface/lerobot/calibration/

For the person on site: do these before leaving, nothing else

Measure four things with a tape measure and photograph them. Floor to table top; floor to the left and right shoulder joint axis (the centre of the pivot at the base of each arm); floor to the head camera lens; robot front edge to the table edge. Also the table length and width. Write the numbers on paper, photograph it, post it in the group.
Mark the four wheel positions on the floor with tape and take one photo from above.
Label the three camera cables and photograph them: which hub port the left wrist, right wrist and head camera are plugged into. Plug them back the same way at the new venue.
Before transport: ask the remote team to release torque, then support the arms and lay them flat, power off, press the head gently down to its lowest tilt, leave the camera cables plugged in.
At the new venue: measure floor-to-table and floor-to-shoulder again and compare with item 1. Table top to shoulder axis should be about 13.5 cm; if it is off by more than 1 cm, put boards under the robot (shim the robot, not the table). Then power on, connect the network; the rest is done remotely.

About 40 minutes. Everything below is for the remote operators; the person on site does not need to read it.

0 · One-paragraph version

The table. All geometry is expressed from the shoulder pivot; the table top sits 0.1349 m below it (left z_min 0.061 − gripper length 0.1959; right 0.065 − 0.1999; measured per arm on 2026-08-24, double-checked, arms agree within 4 mm). At the new venue either adjust the table (or a riser under the robot) so that shoulder pivot to table top = 13.5 cm ± 1 cm, or accept that the learned policies start from zero. The first thing to do on arrival is re-measure with measure_table.py, not trust the old value.

Hard-coded IK chain: 8 config files, about 2 hours, all re-measurable with existing scripts (except the missing table_datum.json and the base scale factor that was never calibrated).

SmolVLA chain: the checkpoints have the old table height, the old head pose (within-episode head std 0.0000 — zero tolerance to head motion) and the old lighting (two batches, G/R 1.28 and 0.98) baked in. Any change is out of distribution; the only fix is re-recording ≥50 episodes and retraining (1–2 days on the 4090). So the goal of the move is not "recalibrate", it is "make the new venue into the old one".

Capture before leaving (section 4): physical dimensions with a tape measure plus photos; preregistered baseline runs in the reference configuration; head reference images and an empty-table plate; a --preview PASS log; all configs packed and committed; both datasets and checkpoints confirmed in the cloud.

1 · What is baked into the venue — grouped by physical dependency

Depends on	File · value	Used by	How to re-measure	Time
Table height (relative to shoulder pivot)	`configs/teleop_safety.yaml` arms.left.z[0]=0.061 / right 0.065; gripper_len_m 0.1959 / 0.1999; home_pose z 0.17 (the boot self-test requires the fingertip ≥100 mm above the table, so a new table height moves home too)	Safety layer every frame (policy_safety.py calls z_floor/y_floor, pitch-compensated), homing, teleop	`measure_table.py --arm left --port …`: torque off, place the fingertip by hand where the cup will stand, FK reads wrist height; once per arm; write into the panel or the YAML by hand. `--pivot` also re-derives gripper length	30 min
	`configs/table_calibrations.json`, newest 20260824-064847, with head pan/tilt and three photos	History and rollback (robot_server calib_restore)	Append a new entry at the new venue, do not edit the old one; the server rejects calibrations where the arms differ >20 mm or spread >10 mm	with above
	`cam_to_arm.py table_datum()` → configs/table_datum.json does not exist; silently falls back to 0.20 m	Head-camera pixel → arm coordinates	No script; one must be written (or note it and move on)	—
	`configs/place_pose.json` (10 demos, 2026-09-07), `pour_pose.json`, `trial_start_pose*.json` (median of each episode's first frame; arm_home.start_poses_for recomputes offline)	Place, pour, trial start pose	Joint angles are absolute and cannot be shifted with the table; place/pour need 10 new demos each; the start pose comes from the dataset and does not change unless the dataset does	place 30 min
	`visual_servo.py` assumes "the cup sits on a fixed plane"; height comes from the approach pose	Visual-servo grasp	Approach pose re-taught with the table height	with teach_grasp
	Stale copy: robot_server.py still quotes gripper length 0.159, right z_min +0.02, table −0.139 — contradicts the YAML	Panel help text	Fix while re-measuring	5 min
Head camera pose	`configs/head_presets.json` cup-grasp-v0 = pan 2003 / tilt 2619 (operator decision 2026-09-04 as the standard view). Note 2619 is 2 ticks above the servo range_max 2617	Head lock in trials, recording, robot_effectors HEAD_PRESET	`head_aim.py --preset cup-grasp-v0`, then look at the image: the five reference frames at /api/reference/head (different cup positions) must all overlay (handbook P-03: "similar brightness does not mean it is looking at the table")	15 min
	`configs/head_reference_{0..4}.jpg`, head_reference_cup_grasp.jpg (2026-08-23)	Head-aim card	These are photographs of Space 21 and mislead the moment you move; re-shoot at the new venue, rename the old ones for the archive	10 min
	cam_to_arm_left/right.json, servo_jacobian_left.json (48 samples, 2026-09-07, head at pan 2000 tilt 2615), wrist_jacobian_left.json, grasp_jacobian.json (2026-09-10), grasp_demos.json (8 demos, 09-10), grasp_tip_offset.json (n=2, std larger than the mean — never usable)	Visual servo, teach_grasp	cam_to_arm docstring: "valid only at the head pose it was fitted at". `teach_grasp.py` full session 12 min (TEACH_GRASP.md); `visual_servo.py --measure-jacobian --go` 10 min; `cam_to_arm.py --sweep` 30 min	50 min
Lighting / colour	`cup_locate.py` CUP_H 12–22, S≥80, V≥75, TABLE_TOP 0.18 (measured 2026-09-10 in daylight on the head camera); cup_mask.py blue/yellow thresholds (09-11); finger_locate.py tape thresholds	Cup locating, cup masking, fingertip locating	Re-shoot 12×12 patches and measure HSV at the new venue; the table must still be white (S<25)	20 min
	Head camera G/R: day 0.95–0.98, night 1.54–1.58; the dataset has two batches, 1.28 / 0.98	SmolVLA input distribution; handbook V-06 reports night results separately	trial_keyframes logs light.gr/lum on every trial; check this number on the first trial at the new venue	0
	Empty-table plate for cup_mask --mask-fill plate	Colour-conditioned picking	Head at cup-grasp-v0, nothing on the table, one 424×240 frame	2 min
USB topology	Camera by-path strings hard-coded in three places (robot_server.py:86, run_policy_trials.py:861, camera_view/identify_cameras): left wrist 2.3.2, right wrist 2.3.3, head 1.1:1.4. The Orbbec already changed port once on 09-03; the two wrist cameras share a serial number and are told apart only by hub port	Recording, trials, previews	Plugging into swapped ports silently swaps left/right. On arrival run the panel's identify_cameras (rotates one wrist, shows which image moves), then `run_policy_trials.py --preview` for camera path/resolution PASS	10 min
USB topology	Serial ports /dev/ttyACM0/1 are identified by motor count, not by port	Every bus script	Nothing to do	0
Floor	base_translate.py CAL_SCALE = 1.00 "fill in after calibrating" — never calibrated; WHEEL_MOUNT_DEG is robot-intrinsic	Base translation (not used in the demo)	Drive 0.5 m, tape-measure the result; carpet and tile differ	10 min
Network	Tailscale 100.107.145.111 travels with the robot; ports 8770/8790/8781/8443/8797/5900 unchanged	Remote panel, upload, VR	Nothing to do	0
Network	fix_hotspot_internet.sh hard-codes WAN_IF=enP8p1s0 (wired, 193.167.0.37), AP_IF=wlP1p1s0, subnet 10.42.0.0/24; iptables do not survive a reboot	Quest hotspot internet	If the venue is Wi-Fi/DHCP, change WAN_IF and re-run; verify with `verify_hotspot_client.sh`	15 min
Robot itself (venue-independent)	Servo calibration ~/.cache/huggingface/lerobot/calibration/robots/xlerobot/xlerobot_arms_only.json (2026-08-16); kinematics L1 0.1159 / L2 0.1350; max_reach_m 0.2459; WHEEL_MOUNT_DEG; pan_sign	Everything	Redo only after a servo swap (lerobot-calibrate, type c at the prompt). The left shoulder_lift stalled at 69 °C on 09-09 and sat above 50 °C all day today; if it is replaced before the move, the whole chain restarts from servo calibration: z_min, gripper travel and pan_sign all re-measured (SETTINGS-AUDIT rule)	swap +1 h

2 · The two chains side by side: what happens when the table height changes

	Hard-coded IK chain (safety layer · arm_home · place/pour · visual servo · teach_grasp)	SmolVLA chain (sep3-clean 54 episodes · left-sd-p0.2-b16-u20000@020000)
Where the table height lives	One number: teleop_safety.yaml z[0]; everything else derives from it	In 13,693 frames of joint angles and images — everywhere
Table +Δ	Not re-measured: the safety floor is computed for the old table, and the policy/servo pushes the fingertip 3 cm into the table before it is stopped (min_wrist_clearance 0.03). Re-measured: the floor is right, place/pour targets must be re-taught	It outputs absolute joint angles for the old table: +1 cm closes 1 cm higher on the cup wall (probably fine); +3 cm the fingertip hits the table and the safety layer HELDs the whole run; −3 cm it closes on air above the rim. No parameter compensates
Head pose changed	cam_to_arm / jacobians all re-fitted (scripts exist)	Zero tolerance: head std 0.0000 across 54 episodes. Wrong tilt equals "facing the wall"; handbook P-03 records a whole batch lost that way
Lighting changed	cup_locate/cup_mask thresholds re-measured (20 min)	Night G/R 1.54 results are reported separately (V-06); a colour temperature absent from training is out of distribution, degree unknown
Cup position range	Covered by teach_grasp demos, can add demos	Five bands in the 120–175 mm mid range (cup-positions.json); left-arm closure pan median −13.5°; outside that is extrapolation
Cost to fix	≈2 h, all on site	Re-record ≥50 episodes + upload + 20k steps on the 4090 ≈ 1–2 days; or leave it alone and change the venue

3 · Adjusting the table height — three cases

The new table is the same height (±1 cm). On arrival: measure_table on both arms → values should fall within 0.061 / 0.065 ± 0.010 → append a table_calibrations entry → head to cup-grasp-v0, confirm by image → identify_cameras + --preview → 5 baseline trials compared with the pre-departure baseline.1 h, policies untouched

Different height, but adjustable (lift table, boards under the table, a riser under the robot). The target is "shoulder pivot to table top = 13.5 cm", not "how tall is the table". Take the pre-departure floor-to-shoulder measurement, subtract the new table height, and that is how much to shim; then follow A. Shimming the robot is better than shimming the table: the table surface, colour and texture stay as the head camera knows them (cup_locate needs a white table).1.5 h

Different height and not adjustable. Re-measure the whole hard-coded chain per section 1 (2 h) and demo with the scripted grasp only (grasp_cup / teach_grasp; 8 demos taught on 09-10, hardware success rate not recorded). SmolVLA can only be re-recorded: on day one record 50 episodes with leader-follower (head at the standard view first, four acceptance checks from the recording guide), /api/dataset/publish on the spot, one night on the 4090.2 days, and the success rate returns to unknown

Where the tolerance comes from: the safety layer keeps 30 mm of wrist clearance above the table, beyond which the policy gets HELD; the demos close mid-wall on a cup about 8.5 cm tall (teleop_config.py:223), so ±1 cm is still on the wall; and the calibration acceptance itself is arms within 20 mm, spread within 10 mm. So ±1 cm means "the policy most likely still works" and ±3 cm means "it certainly does not". The band in between has never been tested — do not bet on it.

4 · What to do before leaving Space 21 — three classes

A is what only the site can give and the repository lacks; B is optional extra runs; C already exists in the repository and only needs backing up. Re-checked on the evening of 2026-09-11: the table-height reading, reference images, plate, lighting patches and all three baselines that were listed under A/B have moved to C.

A · Not in the repository, only doable on site (about 35 min)

#	What	Tool	Where it goes	Time
A1	Tape measure: floor→table top; floor→left and right shoulder pivot; table top→shoulder pivot (should be ≈13.5 cm; cross-check against the YAML); floor→head camera lens; robot front edge→table edge; table length and width	Tape measure + phone photos with a scale	configs/venue-space21-2026-09-11.json + photos	20 min
A2	Parking spot: tape marks for the four wheels, one top-down photo	Tape, phone	same	5 min
A4	--preview PASS log (camera paths / resolutions / state dims): run `run_policy_trials.py --preview` when the machine is idle and save to 05-training/trials/preview-space21.log; not yet run on the evening of 09-11 because teach_present held the bus	train310 python	05-training/trials	5 min
A3	Which hub port each camera cable is in: label and photograph (left wrist 2.3.2, right wrist 2.3.3, head 1.1). The device-side topology is already saved: 05-training/trials/usb-topology-space21-2026-09-11.txt (lsusb -t + by-path)	Labels, phone	04-reports photos	10 min

B · Optional extra runs (same-configuration results already exist; more only adds sample size)

#	What	Already have	Time
B1	Right arm, 5 more trials at P2–P4, today's recipe	09-11 6/6 (trials-20260911-152009.json)	25 min
B2	Left arm, 5 more trials, cup in the left half of the frame, near	09-09 15/20 (28 runs), 09-11 early morning 4/4	20 min
B3	A night set of head reference frames (the day set exists)	early-morning trial keyframes serve as night references	5 min

C · Already in the repository — only confirm it is backed up (about 20 min)

What exists	Where	The only action
Table calibration and safety envelope	configs/table_calibrations.json (20260824-064847, per arm, 4 mm apart); configs/teleop_safety.yaml	One tar into ~/incoming and download it; git commits are the operator's (20+ untracked files). 10 min
Head presets, reference images, empty-table plate	configs/head_presets.json; head_reference_{0..4}.jpg + head_reference_cup_grasp.jpg (08-23); head_plate_cup-grasp-v0.jpg (09-11); every init keyframe from today's trials
Lighting patches and night data	cup_locate.py comments (HSV measured 09-10); trial_keyframes logs light.gr per trial (day 0.95, night 1.54)
Baselines	left 09-09 15/20, right 09-11 6/6, left at night 09-11 02:xx 4/4; all in trials-*.json and STATUS.md
Camera→arm calibration, jacobians, demos	configs/cam_to_arm_*.json, servo_jacobian_left.json, wrist_jacobian_left.json; scripts/grasp_jacobian.json, grasp_demos.json, grasp_tip_offset.json
Scripted poses	configs/place_pose.json, pour_pose.json, trial_start_pose*.json, marker_offset.json, pan_sign_measured.json, pan_range_measured.json
Servo calibration	~/.cache/huggingface/lerobot/calibration/ (robot-intrinsic, venue-independent)
Cup positions, trial records, USB topology	05-training/trials/cup-positions.json, traces/, keyframes/, usb-topology-space21-2026-09-11.txt
Datasets	sep3-clean (54 ep), 0826sep8-full (73 ep)	Check meta/upload.json says ok (R-05). 5 min
Models	05-training/…/checkpoints/020000 (left and right)	Confirm the same-named repos exist under HF Suyang99. 2 min

Totals: A about 40 min and it is the hard requirement — without it you do not know how much to shim at the new venue; B optional; C about 20 min.

5 · Order of work on arrival (about 2 h + baseline)

Before powering on: tape-measure floor→table and floor→shoulder, compare with A1, decide whether to shim (section 3).
check_arms.py: all 17 servos answer, temperatures normal (left shoulder <50 °C before running).
identify_cameras → --preview PASS; if it does not match, re-plug and re-run — do not edit the by-path strings in code.
measure_table.py on both arms → append table_calibrations → update YAML z[0] (and home z if the table moved).
head_aim --preset cup-grasp-v0 → overlay the five reference frames → shoot a new plate.
arm_home both arms, self-test passes (P-01).
Lighting: read light.gr on the first trial; above 1.3 report as night; re-measure cup_locate/cup_mask thresholds if needed.
teach_grasp.py 12-minute re-teach (scripted grasp line); place_pose re-teach if the demo places cups.
Baseline 5 trials, six configuration items identical to section 4 B; write them next to the pre-departure baseline in STATUS.
Quest hotspot: set WAN_IF to the venue's interface, verify_hotspot_client.sh.

6 · Known gaps (not solved on this page)

No transport document: power-down order, torque release, wheel locks, cable restraint. Suggestion: check_arms --release → power off → lower the head to minimum tilt by hand → leave camera cables plugged in (unplugging means redoing A5).
E-stop does nothing while leader-follower is running (HANDOFF-2026-09-09-NIGHT): the bug is still there when recording at the new venue.
Left shoulder_lift: stalled at 69 °C on 09-09, above 50 °C all day today with 6–7° under-torque: decide whether to replace it before the move; if replaced, follow the "robot itself" row for the full recalibration.
table_datum.json missing, CAL_SCALE never calibrated, tilt 2619 two ticks over range_max: three old issues; the move is the moment to clear them.
All durations on this page are estimates (evidence level: not run); script existence and parameters were checked (evidence level: ls/grep run).

Values sourced per row; table −0.1349 m and z_min from teleop_safety.yaml and table_calibrations.json 20260824-064847; written 2026-09-11 afternoon with the robot in Space 21.