Co-Creation in Bloom
From Robot Intelligence to Human–Robot Intelligence
机器人不会仅仅因为能够操作物体，就自然成为我们日常生活的一部分。它们还需要能够自然地与人协作。在这个项目中，我们通过艺术创作探索人机协作。艺术只是媒介。我们的目标，是构建能够理解人、与人协调，并和人共同创作的机器人。
Robots won't become part of our daily lives simply because they can manipulate objects. They need to collaborate with humans naturally. In this project, we explore human-robot collaboration through artistic creation. Art is only the medium. Our goal is to build robots that understand, coordinate, and create together with people.
Introduction · 引言
当机器人拿起画笔
在具身智能的研究中，我们习惯用 success rate 衡量一个 manipulation 任务的好坏：抓没抓起来、放没放对位置、整段轨迹是否完成。但当任务的对象不再是工业件，而是一片片轻巧、不可重复的花瓣时，"成功"这个词就变得模糊起来 —— 一幅画没有标准答案，只有一种在共同创作中逐渐生长出来的形状。
我们设计了一个非常简单的设定：把一桌花瓣作为共享材料，把一张黑色画布作为共享画面，把机器人和人放在同一张桌子旁边。两者轮流"落笔"，每一次"落笔"都建立在对方刚刚留下的痕迹之上。这就是 Artistic Co-Creation Manipulation。
Introduction
When the Robot Holds a Brush
In embodied manipulation research we are used to measuring tasks by success rate — did the gripper close, did the object land in the bin, did the trajectory complete. But when the medium is a pile of light, non-repeatable flower petals, the word "success" loses its sharp edge. A painting has no ground-truth answer; it has a shape that grows out of two agents working together.
The setup is intentionally minimal. A table of petals as the shared material, a black canvas as the shared surface, and a robot and a human seated at the same table. They take turns. Every stroke is placed on top of the trace the other one just left. We call this Artistic Co-Creation Manipulation.
Chapter 1 · 交互循环
人 — 机器人 — 人：一个不断闭合的回路
所有等待与犹豫都被完整保留。我们想呈现的不是一个精心剪辑的 demo，而是双向协作真实发生的节奏。
人先落下几片花瓣，给出初始的方向；机器人观察人留下的痕迹，拾起新的花瓣作出回应 —— 延续这个构图，或有意打破它；人再介入，在机器人留下的形状上继续生长。作品不属于任何一方，而是在彼此让步与回应中逐渐显形。
一镜到底 · 1 倍速 · 点击右上角的「Play full video」可在全屏查看完整视频。
Chapter 1 · Interaction Loop
The Human–Robot Interaction Manipulation Loop
Every pause and hesitation is preserved. What we wanted to show is not a polished demo reel, but the real rhythm of two agents collaborating.
The human places a few petals, sketching a direction. The robot reads the trace and answers with petals of its own — carrying the composition forward, or deliberately breaking it. The human builds on what the robot left. The piece belongs to neither side; it emerges through reciprocal yielding and answering.
One continuous take · 1× playback · Click "Play full video" above for the full screen view.
Chapter 2 · Demo · 四个阶段
从自主到共创的四幕
Chapter 2 · Demo · Four Stages
Four Acts, from Solo to Duet
自主创作 · Autonomous Creation
无人在场时，机器人持续、不停顿地独自创作：从花堆中拾取轻巧的花瓣与枝条，精准而顺滑地逐一摆放到黑色画布上。
Autonomous Creation
With no one at the table, the robot creates alone — continuously, without pausing — picking delicate petals and branches from the tray and placing them onto the black canvas with precise, smooth grasps.
察觉与让位 · Notice & Yield
人走进视野并坐到桌旁。没有任何开关或指令，同一个策略仅凭视觉观察便察觉到人的到场，停下当前动作，让出画布，进入共创状态。
Notice & Yield
A person steps into view and sits down. With no switch, button, or mode flag, the same policy notices the human presence purely from visual observation — it pauses mid-motion and yields the canvas, shifting into co-creation.
轮流交互共创 · Turn-Taking Co-Creation
人放置 → 机器人观察 → 机器人放置 → 机器人等待 → 人放置……两者按人的节奏交替"落笔"，在彼此留下的痕迹上继续生长这幅作品。
Turn-Taking Co-Creation
Human places → robot observes → robot places → robot waits → human places… The two alternate at a human-paced tempo, each stroke building on the trace the other just left.
察觉与恢复 · Adapt & Restore
人刻意移走了几朵花。机器人察觉构图被改变，在线调整下一步动作：回到花堆重新拾取，并将花朵一一放回原先的位置。
Adapt & Restore
The person deliberately removes a few flowers. The robot notices the altered composition and adapts its next move online — returning to the tray and restoring each flower to its original position.
Chapter 3 · 我们的 Focus
不止 success rate
在 Artistic Co-Creation Manipulation 里，单一的成功率讲不完这个故事。我们用四个指标来评估：
Success Rate · 成功率
每一次抓取与摆放是否可靠落地？我们统计单次摆放、完整会话，以及共创期间的成功比例——这是所有上层行为的基本盘。
Rhythm · 节拍一致性
机器人是否跟得上人的节奏？我们用机器人与人的平均取放耗时之比来衡量——比值接近 1，意味着同频推进，而不是拖在后面。
Adaptation · 对介入的适应
当人重新排布构图后，机器人是响应当前画布，还是照搬示范的顺序？我们统计其下一朵花落在正确区域的比例。
Completion · 自主知止
当画面到达完成态，机器人能否自行收手？我们统计无需人工干预、正确自主终止的会话比例。
这四个指标合在一起回答同一个问题：我们要衡量的不止是"抓没抓稳"，更是人和机器人之间的关系。
Chapter 3 · Our Focus
Beyond Success Rate
A single success rate doesn't tell the whole story in Artistic Co-Creation Manipulation. We evaluate along four measures:
Success Rate
Does each pick-and-place land reliably? We count per-placement success, full sessions, and the rate during co-creation — the base layer under everything else.
Rhythm
Does the robot keep pace with its partner? We take the ratio of the robot's average pick-and-place time to the human's — near parity means keeping up, not lagging behind.
Adaptation
After the human rearranges the layout, does the robot respond to the current canvas or replay the demonstrated order? We measure how often the next petal lands in the correct region.
Completion
When the canvas reaches its finished state, does the robot stop on its own? We measure the share of sessions that terminate correctly without intervention.
Together, the four answer one question: not just whether each grasp landed, but the relationship between human and robot.
Chapter 4 · 数据构成
数据金字塔
Chapter 4 · Data Composition
The Data Pyramid
顶层 · Rollout
策略在真机上执行（rollout）时产生的数据。
Top · Rollout
Data generated as the policy runs on the real robot.
中层 · 人机交互共创
与人轮流交互摆花 200 条，由多位协作者轮换采集；花朵种类有泛化，摆放位置的泛化幅度较小。
Middle · Human–Robot Co-Creation
200 turn-taking interaction episodes, collected with rotating human partners; flower species are varied, with smaller variation in placement positions.
底层 · 自主任务遥操
机器人自主摆花的遥操演示共 525 条，固定 setup 与泛化 setup 混合；花朵种类与摆放位置各有泛化，并定期更换花材。
Base · Autonomous Teleoperation
525 teleoperated episodes of autonomous arrangement, mixing fixed and randomized setups; both flower species and placement positions are varied, with flowers rotated periodically.
Chapter 5 · 到访与联系
来现场，或与团队聊聊
Chapter 5 · Visit & Connect
Visit the Studio or Meet the Team
来现场一起创作
想亲手体验人机共创，可以预约来我们的工作空间完成一幅花瓣画。
Create With Us in Person
Book a visit to experience human-robot co-creation and make a petal painting with the robot.
提交后，我们会通过邮件确认到访时间。 We'll confirm your visit by email.
预约已收到
我们会尽快通过邮件与你确认到访时间。期待现场见！
Request Received
We'll email you soon to confirm a time. See you at the studio!
项目团队
如需研究交流或项目合作，欢迎直接联系我们。
Project Team
For research discussions and project collaborations, contact us directly.
Appendix · 更多演示
More Demos
Appendix · More Demos