Sim2Real leaps: OmniReset
Research updates show OmniReset scaling RL can achieve zero-shot sim-to-real dexterity, while robotics 'world model' work is stitching vision, physics, and control into more generalist motion models. Together these posts argue the field is closing the gap between simulated training and real-world manipulation. (x.com) (x.com) (x.com)
OmniReset’s author list names Patrick Yin and Tyler Westenbroek as equal contributors and includes researchers from University of Washington, NVIDIA, and Microsoft Research; the paper appeared on arXiv in March 2026 and is listed as an ICLR 2026 poster. (arxiv.org) The project’s pipeline programmatically generates reset states and then runs large-scale state-based RL — the authors report training at scale across 64K+ parallel simulated environments before distilling policies to RGB visuomotor controllers. (weirdlabuw.github.io) OmniReset’s experiments show emergent multi-phase, contact-rich behaviors (examples include pushing and flipping a drawer and picking-and-twisting a table leg) and claim coverage of initial conditions more than 300× wider than the strongest baselines. (arxiv.org) DreamDojo, led by NVIDIA GEAR Lab and academic partners, was pretrained on a 44,000+ hour egocentric human-video corpus (DreamDojo‑HV) that the team reports contains ~6,015 unique tasks, 1M+ trajectories, 9,869 scenes and 43,237 objects. (arxiv.org) The DreamDojo release includes two model variants (reported 2B and 14B parameter sizes), a two‑stage workflow of human-video pretraining plus robot-specific post‑training, and a distilled real‑time rollout target of roughly 10 frames per second for long-horizon generation. (github.com) Taken together, the two threads follow complementary scaling paths: OmniReset trades large simulated reset diversity and PPO-scale RL to produce task trajectories for distillation, while DreamDojo trades human-video scale and learned latent actions to build pixel‑based world models that can be post‑trained for robot control; both projects are published with code and benchmarks to enable follow‑on post‑training and evaluation. (openreview.net)
