이미지: METAL LAB 생성
Summary
- Apple has introduced MoMo, which treats how a robot manipulation motion is executed—its speed and intensity—as a single learnable variable
- In experiments on six real-robot tasks, stable, dynamic, and intermediate motion modes were clearly distinguished by joint velocity, acceleration, and approach angle
- Even for tasks demonstrated in only one style, the model maintained task success rates while generalizing to new motion modes, confirming its generalization ability
- 발표
- 애플 머신러닝 리서치, 2026년 7월 논문 공개
- 저자
- Yuhan Hu, Hugues Thomas 등 6명
- 구조
- 시공간 행동 토크나이저 + 행동복제 트랜스포머로 이뤄진 2단계 모방학습
- 실험 대상
- 실제 로봇 조작 과제 6종
- 확인된 차이
- 관절 속도·가속도·엔드이펙터 접근 각도에서 동작 모드별 차이 관측
Robots need to tell "fast" apart from "gentle" too
Apple has unveiled a framework called "MoMo" that trains a robot arm to adjust the speed and intensity of its movements to fit the situation, even when performing the same task. The paper, titled "MoMo: Dial Motion Mode in Robot Manipulation with Spatiotemporal Action Tokenization," was published on Apple Machine Learning Research in July.
MoMo is a two-stage imitation learning framework—a training method in which a robot learns behavior by watching videos of human demonstrations. First, a spatiotemporal action tokenizer converts the robot's motion into tokens that encode both time and space information. Then, a behavior-cloning transformer generates the actual movement based on these tokens, taking as input both the task type and a continuous conditioning value called "motion mode." In experiments on six manipulation tasks with real robots, varying this motion mode value produced stable motions, dynamic motions, and motions somewhere in between, and human evaluators were able to tell the difference. Actual differences were confirmed in joint velocity, acceleration, and the approach angle of the fingertip when grasping an object.
Why "how" it moves matters
Until now, robot manipulation research has largely focused on "what" to do. The key question was whether the robot could accurately perform a given task, such as picking up a cup or opening a drawer. But humans perform the same action differently depending on the situation. A glass is carried slowly and carefully, while a sturdy plastic container is grabbed quickly and moved almost by tossing it. This "difference in execution style" has often been treated as noise and ignored in robot learning.
This is exactly where MoMo focuses its attention. The idea is that if motion mode is treated as a separate, reusable component detached from the task itself, the "carefulness" learned in one task can be transferred directly to an entirely different task. The researchers reported that even for tasks demonstrated in only one style, the robot accepted requests for undemonstrated new motion modes without significantly losing task success rate. This was presented as evidence that the robot can generalize to task-mode combinations it has never been trained on.
In the field of robot learning, there has been a growing trend toward foundation-model approaches that train on large volumes of human demonstration video to adapt to diverse objects and environments. Dyna Robotics' release of a robot model trained on one million hours of human video footage is part of the same trend. MoMo stands apart from this large-scale training approach in that, with relatively little data, it makes "style" of motion a new, controllable axis.
So what changes
Until now, telling a robot to "be more careful" required either collecting an entirely new set of demonstration data or manually writing rule-based control logic. If the MoMo approach becomes practical, a single trained motion-mode dial could allow speed and intensity to be adjusted on the fly across multiple tasks. Potential applications include a single arm in a warehouse handling fragile and sturdy items differently, or a collaborative robot working near humans slowing its movements depending on the situation. However, since this research was validated on only six tasks in a specific lab environment, further validation will likely be needed before it can be applied to commercial robots.
