Real2Sim2Real presents a promising opportunity for zero-shot robot deployment 🤖, but how does this paradigm translate to tricky, contact-rich applications such as the game of billiards 🎱?
We argue that an agent can learn the underlying dynamics parameters through interaction!
We argue that an agent can learn the underlying dynamics parameters through interaction!
Can an agent learn a pool table’s physics by playing on it? 🎱
While agents can close the visual gap, the difference between real world and sim dynamics remains significant. For CoRL late breaking🤠, we let an agent take real shots to try and close this real2sim dynamics gap: cue-dynamics.github.io
Very fun short project with @XilunZhangXZ @MatthewFoutter @Luis_A_Pabon @drmapavoneView quoted post on X
