Omnireset paper excerpt: generating robot resets from grasp sampling and near-goal offsets

YouJiacheng · x · 2026-11-19

In a follow-up to the reset-design debate, YouJiacheng shares an Omnireset paper excerpt detailing its reset generation: sample 1000 feasible grasp points on the target object using Mittal et al. (2023), then apply small random forces to dislodge the object from the goal, producing a continuum of configurations like partially-inserted peg-in-hole states. This yields reset classes including Stable Grasp Resets (random aerial object spawn with gripper at a feasible grasp) and Near-Goal Resets (dense coverage of contact-rich near-goal states) — a concrete answer to how reset buffers can be systematically generated.

Related event: Reset Buffer Design Remains an Overlooked Challenge in Robot RL(2 posts)→

Original post →

More from Embodied

Embodied channel →