OmniReset scales zero‑shot sim2real for dexterity
OmniReset published work showing RL at scale can achieve zero‑shot sim‑to‑real dexterous manipulation by prioritizing state coverage over demo collection—reframing the data tradeoffs in dexterous RL. The approach suggests broad state exploration in sim can replace costly real demonstrations for some tasks. (x.com)
Emergent Dexterity via Diverse Resets lists Patrick Yin and Tyler Westenbroek (equal contribution) among lead authors and shows affiliations with University of Washington, NVIDIA, and Microsoft Research on the arXiv submission. (arxiv.org) The paper frames its method around on‑policy reinforcement learning trained with a single reward function, fixed algorithm hyperparameters, no curricula, and no human demonstrations while programmatically generating diverse simulator reset states to expand state coverage. (arxiv.org) OpenReview and the ICLR poster abstract record that the authors included seven distinct long‑horizon, contact‑rich manipulation tasks and claim OmniReset scales with compute to solve problems beyond prior techniques’ reach. (openreview.net) The manuscript reports distilling the large on‑policy policies into RGB visuomotor policies that transfer zero‑shot to real robots, with emergent retrying behaviors and “substantially higher success rates” than baseline methods in their real‑world tests. (arxiv.org) The authors released code and assets on GitHub with a project release dated February 18, 2026, and an interactive demo page (weirdlab/OmniReset) that visualizes the reset procedure and distilled policy behaviors. (github.com) OpenReview shows the paper was accepted as an ICLR poster (decision posted January 26, 2026), and the project webpage linked in the paper provides downloadable models and experiment details for replication. (openreview.net)
