OmniReset

OmniReset#

Paper: Emergent Dexterity via Diverse Resets and Large-Scale Reinforcement Learning (ICLR 2026)

Project website: omnireset.github.io

Full Pipeline#

The full OmniReset pipeline from custom task creation to real-robot deployment:

1. Create New Task
assets & variants

→

2. Train RL Policy
resets & training
★ most users start here

→

3. Sys-ID & Finetune
sim2real alignment

→

4. Distill & Deploy
vision policy & real robot

Tip

Most users only need step 2. If you’re training on one of our 6 existing tasks, jump straight to Collect Resets & Train RL Policy.

Create a New Task – Prepare USD assets, register object variants, verify in sim.
Collect Resets & Train RL Policy – Collect reset states and train an RL policy from scratch. Start here for most use cases.
Sim2Real: SysID & RL Finetuning – Robot calibration & USD, system identification, camera calibration, then ADR finetuning.
Distillation & Deployment – Collect demos and train your own ResNet18-MLP vision policy. Deploy on real robot.

Pipeline

Compute & Hardware Requirements#

Stage	Requirements
Policy evaluation	1 GPU.
RL training	4 GPUs, 24+ GB VRAM each (e.g. L40S, 4090). Cube/Peg converge in ~8 hours on 4x L40S.
RL finetuning	1–4 GPUs depending on task (see Sim2Real: SysID & RL Finetuning for per-task env counts). Peg converges in ~8 hours on 1x L40S.
Demo collection	1 RTX GPU, 24+ GB VRAM (32 envs fit on an RTX 4090). 10K demos ~2 hours.
Vision policy training	1 GPU. ~2 days of training on a H200 for transfer. ~1 day of training on a H200 for sim-only distillation.
Real-robot deploy	UR5e/UR7e + Robotiq 2F-85 + 3x Intel RealSense (D415/D435/D455).

BibTeX#

@inproceedings{
   yin2026omnireset,
   title={Emergent Dexterity via Diverse Resets and Large-Scale Reinforcement Learning},
   author={Patrick Yin and Tyler Westenbroek and Zhengyu Zhang and Joshua Tran and Ignacio Dagnino and Eeshani Shilamkar and Numfor Mbiziwo-Tiapo and Simran Bagaria and Xinlei Liu and Galen Mullins and Andrey Kolobov and Abhishek Gupta},
   booktitle={The Fourteenth International Conference on Learning Representations},
   year={2026},
   url={https://arxiv.org/abs/2603.15789}
}

OmniReset

Contents

OmniReset#

Quick Start (Try in 2 Minutes)#

Full Pipeline#

Compute & Hardware Requirements#

BibTeX#