多种方式搬箱子 · 动作捕捉Carrying a Box in Many Ways · Motion Capture

我们在训练人形机器人：看人示范一种搬箱子的方式，就照这种方式搬。这一页说明需要拍什么、怎么拍。We are training a humanoid robot to watch a person carry a box one way and then carry it the same way. This page explains what to record and how.

什么样的数据对我们的研究最有用What makes this data useful for our research

我们收到数据后，主要看这三件事;每一件后面列的是我们会实际检查的指标。When the data arrives we look at these three things; under each are the measurements we actually check.

1

动作干净Clean capture

机器人要在物理仿真里一帧一帧地复现这段动作。脚在地上打滑、箱子瞬移、手穿进箱子，在物理上都做不到，机器人就学不会这一段。The robot replays the motion frame by frame in a physics simulator. Sliding feet, a jumping box or hands passing into the box are physically impossible, so the robot cannot learn that take.

我们检查WE CHECK
脚着地时滑了多远foot slide while in contact箱子标记点丢了几帧frames with lost box markers手和箱子表面贴得多近hand-to-box surface distance
2

搬法多样Many different ways

机器人看一段示范，就照那种方式搬。同一种搬法拍很多遍只算一个例子;不同的搬法越多，它能学会照着做的就越多。The robot watches one demonstration and carries the box that way. Many takes of one style count as one example; every new style adds something it can learn to imitate.

我们检查WE CHECK
明显不同的搬法有多少种number of clearly different styles同一种搬法里快慢、高低、情绪的变化variation within a style: speed, height, mood
3

方便拼接Easy to stitch

拐弯、走到别的地方，我们用动作生成模型补出来。它需要每条的开头和结尾都是稳定的站姿，中间走直线。We fill in turns and other destinations with a motion-generation model. It needs every take to begin and end in a still stand, with a straight walk in between.

我们检查WE CHECK
首尾各站稳多久length of the still stand at both ends箱子是否从地上起、回到地上box starts and ends on the floor走的是不是直线how straight the walk is

怎么拍：两段 30 秒的视频How to record: two 30-second videos

一条完整的录制One full take站稳 → 拿起 → 直走 3 m → 放下 → 站稳stand → pick up → walk 3 m → put down → stand
每种搬法拍哪几条Which takes for each style完整 2 条，只拿起 1 条，只放下 1 条full ×2, pick-up only ×1, put-down only ×1

我们现有的低质量版本Our current low-quality versions

我们已经有一些低质量的动作，下面这些只是为了说明搬法长什么样;更多可以拍的搬法列在后面。We have some low-quality motions already; these only show what the styles look like. More styles to record are listed after them.

A1 双手托底端在身前Front carry
A2 抱在胸前Hug to chest
A3 右腋下夹Under right arm
A4 左腋下夹Under left arm
A5 扛在右肩On right shoulder
A6 扛在左肩On left shoulder
A7 掌心向上平托Palms-up tray
A8 背在身后Behind the back

想拍的搬法(按优先级)Styles to record, by priority

带视频或链接的是网上找到的类似动作，只作参考;搬的东西不同没关系。Clips and links show similar motions found online, for reference only; the object being carried does not matter.

第一优先First priority最想要

手要换好几次位置的搬法，以及我们完全没有的搬法，最需要真人动捕。Styles where the hands change position several times, and styles we do not have at all, need real mocap the most.

A5 / A6
扛在右肩 / 左肩On the right / left shoulder
从地上拿起，甩上肩，一只手扶着走，到点再放回地上lift from the floor, swing it onto the shoulder, steady it with one hand, then put it back on the floor看网上示范 ↗Watch online ↗
A3 / A4
夹在右腋下 / 左腋下Under the right / left arm
双手拿起后夹到一侧腋下，单手扶着走pick it up with both hands, tuck it under one arm, walk with one hand on it
A8
背在身后Behind the back
拿起后转到身后，双手在背后托住after lifting, move it behind the back and hold it there with both hands看网上示范 ↗Watch online ↗
A7
掌心向上平托Palms-up tray
两前臂端平，箱子放在掌心和前臂上forearms level, the box resting on palms and forearms
A10 / A11
顶在头上 / 举过头顶On the head / overhead
箱子顶在头上双手扶;或双臂伸直举在头上方balance it on the head with both hands, or hold it overhead with straight arms看网上示范 ↗Watch online ↗
D1–D4
走到一半换搬法Switch style mid-walk
例如：身前端 → 右肩;左腋下 → 右腋下;抱胸 → 单手侧提;其他组合也可以for example front → right shoulder, left arm → right arm, hug → one hand at the side, or any other switch
C1–C6
同一个箱子，演出不同属性Same box, acted as if different
像很重、像易碎、像很烫、像很脏、像装满水、急匆匆as if very heavy, fragile, hot, dirty, full of water, or in a hurry看网上示范 ↗Watch online ↗
第二优先Second priority有时间再拍if time allows

两只手全程不换位置的搬法，以及不同的拿起、放下方式。Styles where the hands stay in one place, plus different ways to pick up and put down.

A9
双手侧面夹Side squeeze
双手按住箱子左右两侧，手臂伸直在身前palms press the two sides, arms straight in front
A13
单手侧提One hand at the side
一只手抓住箱子边缘，提在身体一侧one hand grips the edge, box hanging at the side
A12
垂手低提Low in front
手臂向下伸直，箱子在大腿前arms hanging straight, box in front of the thighs
B1–B3
不同的拿起方式Different ways to pick up
深蹲直背、弯腰、单膝跪地squat with a straight back, bend at the waist, kneel on one knee看网上示范 ↗Watch online ↗
B4–B6
不同的放下方式Different ways to put down
慢慢放稳、离地 20 cm 内松手(安全前提下)、单膝跪地放set it down slowly, release from within 20 cm of the floor (only if safe), kneel to place it
可以少拍Fewer takes needed已有可用的版本we have usable versions

这两种各拍 1 条完整的即可。One full take of each is enough.

A1
双手托底端在身前Front carry
双手在箱底，箱子在腰前both hands under the box, box at the waist in front
A2
抱在胸前Hug to chest
两臂环抱，箱子贴胸arms wrapped around, box against the chest
清单之外的搬法也很欢迎：越不一样越好，请自由发挥。Styles beyond this list are very welcome: the more different, the better.

其他细节(需要时点开)Other details (open when needed)

场地、箱子、演员Floor, box, performer
  • 地上贴 START、A、B 三个标记;箱子放在 A 点地上，A 到 B 3 米直线。
  • Tape START, A and B on the floor; the box sits on the floor at A, and A to B is 3 m in a straight line.
  • 箱子：约 25 × 15 × 20 cm 的纸箱，内加配重到约 1 kg;六个面都贴标记点，抱住时也能被看到。
  • Box: a cardboard box of about 25 × 15 × 20 cm, weighted to about 1 kg; markers on all six faces so it stays tracked when hugged.
  • 箱子长边朝向演员，A、B 两处朝向一样。
  • The long side of the box faces the performer, the same at A and B.
  • 演员一位，身高 1.60–1.80 m 左右;开拍前录一条 T 姿态，并记下身高。
  • One performer, roughly 1.60–1.80 m; record one T-pose take first and note the performer's height.
  • 采样率 120 Hz 左右即可;手指可选。
  • About 120 Hz is great; fingers are optional.
交付格式Delivery
  • 人体：每条一个 FBX 或 BVH;请注明单位和哪个轴朝上。
  • Body: one FBX or BVH per take; please note the units and the up axis.
  • 箱子：同一 FBX 里的一个刚体，或一个 CSV(time, x, y, z, qx, qy, qz, qw),与人体同一坐标、同一时钟。
  • Box: a rigid body in the same FBX, or a CSV (time, x, y, z, qx, qy, qz, qw) in the same frame and clock as the body.
  • 文件名里带上搬法编号(如 A5、D1)即可。
  • Put the style ID (such as A5 or D1) in the file name.
现场小清单On-set checklist
  • START、A、B 已贴好，A 到 B 约 3 m 直线
  • START, A, B taped; A to B about 3 m straight
  • 箱子被抱住时也能稳定追踪
  • The box stays tracked even when hugged
  • 每条开头、结尾都站稳
  • Every take starts and ends with a still stand
  • 箱子从地上拿起、放回地上
  • The box goes floor to floor