Alex Open Research Wiki

OmniReset

2 min read

OmniReset

Summary

OmniReset is an ICLR 2026 dexterous-manipulation framework that programmatically generates broad simulator reset distributions, trains state-based policies with large-batch PPO, and distills those policies into RGB visuomotor controllers for zero-shot real-robot deployment.

Its central mechanism is broad initial-state coverage rather than an explicit ordered curriculum: near-goal, grasped, near-object, and reaching resets expose recurring robot-object interactions, and PPO’s competence expands backward from shorter to longer-horizon starts.

Role In The Wiki

OmniReset is a curriculum/data-generation analogue, not a JEPA or learned world model. It contributes a concrete answer to the question which state neighborhoods must training expose before long-horizon behavior becomes learnable?

The transferable interface for latent-state time-series modeling is:

valid state or trajectory coverage proposal
  -> context/target starts at several horizons
  -> current learner identifies solvable local transitions
  -> competence frontier expands while support is retained

For offline multivariate time series, reset means resampling valid observed windows unless a trusted simulator exists. Generated start states must preserve temporal reachability, exogenous variables, event history, and action/control-input semantics.

Evidence

Official Artifacts

Artifact caveat: the public code/checkpoints were not executed during this ingest. The UWLab repository is BSD-3-Clause, but checkpoint/data artifact terms should still be checked at the exact asset level before reuse.

Claim Boundary

The project page’s no reward shaping shorthand is broader than the paper. The paper uses fixed task-agnostic dense reach, distance, and smoothness terms plus sparse success and safety terms. It also uses an action-space-reduction curriculum during sim-to-real finetuning. The defensible claim is no task-specific reward tuning or explicit ordered curriculum for the simulation behavior-generation experiments.

Relation To Foundation TSFM Agenda

Use the source-level agenda mapping in omnireset-2026 rather than duplicating verdict rows here.

At the entity level, OmniReset is physical-robotics evidence that state-distribution design can determine whether long-horizon learning receives useful signal. It is adjacent to foundation time-series curriculum design, not direct evidence for JEPA, numeric time-series representations, or action-conditioned learned dynamics.

Graph View

Backlinks