MoMo lets robots dial motion modes for adaptive manipulation
Robots can now vary their execution style like turning a dial.
MoMo tackles a key limitation in robot manipulation: robots must not only perform tasks accurately but also adapt how their actions unfold depending on context. The researchers at the MoMo team (Yuhan Hu et al.) propose a two-stage imitation-learning framework. First, a spatiotemporal action tokenizer breaks down robot actions into discrete tokens that capture both spatial and temporal variations. Then, a behavior-cloning transformer takes the task instruction plus a continuous motion-mode condition (a dial-like parameter) and predicts the appropriate token sequence. This allows the robot to smoothly vary its execution style—from steady and precise to dynamic and fast—just by adjusting the dial.
In experiments across six real-robot manipulation tasks, MoMo generated distinct motion modes that human raters could reliably classify. The modes differed not only in qualitative appearance but also in measurable metrics: joint speed, acceleration, and end-effector approach pitch. Crucially, when a task was only demonstrated in a single mode, MoMo could still transfer a user-requested mode to that task, preserving task success. This demonstrates compositional generalization—the robot can combine an unseen task with a known motion mode, effectively reusing behavioral factors across tasks. The work opens the door to more expressive and adaptable robot assistants.
- MoMo uses a spatiotemporal action tokenizer and behavior-cloning transformer to condition on a continuous motion-mode dial.
- Across six real-robot tasks, varying the dial produced steady, dynamic, and intermediate behaviors distinguishable by human raters.
- MoMo transfers unseen motion modes to new tasks while preserving task success, showing compositional generalization.
Why It Matters
Robots can now fluidly adjust their motion style across tasks, making them safer and more intuitive for human collaboration.