In plain words
The paper introduces DONIC, a robot method that uses one controller to follow motion references involving two hands and an object. It also presents BiHOI, a collection of recorded two-hand interactions turned into robot references.
In simulation, the same DONIC controller tracked 86.79% of unseen trajectories involving training objects and 57.32% of clips with new objects. A trajectory or clip counted as successful only when both hands met the object-and-hand tracking requirements.
The research
How can one robot controller follow a two-hand motion reference when the hands take different roles and touch the object at different times? Hand-pose accuracy alone does not ensure the object moves as required, and fully tracked examples with sustained two-hand contact are scarce.
DONIC trains one robot controller with object-motion and per-hand contact feedback, plus verified demonstrations, to track diverse two-hand references.
The paper asks how one robot controller can track two-hand references when hand roles and contact timing change, and hand-pose accuracy alone may not ensure the object follows its motion reference. Fully tracked examples with sustained two-hand contact are also scarce.
DONIC trains one shared controller using object-motion and per-hand contact feedback, then compares it with two adapted methods on held-out simulation splits.
Without further training, it met both-hands tracking thresholds on 86.79% of unseen trajectories involving training objects and 57.32% of clips with held-out objects, supporting tracking across these tested references.
These results support reference tracking in the tested simulations. They do not show that DONIC can perform unconstrained tasks or generalize to new operation categories.
Analysis
Deep
- Original paper authors
- Mingze Sun, Ziang Cao, Yuan Fei, Qiankuan Tao, Zihan Wang
- Submitted to arXiv
- License
- CC BY 4.0