Latest Papers - Heungwoo/research GitHub Wiki
π Latest Papers β preprint tracker
Recent preprints and notable model releases reviewed ahead of (or without) venue publication. These are fast-moving, often v1/early-revision works: numbers are quoted from the paper but unreplicated, and venue-of-record is pending. Once a paper is accepted, its page is cross-linked from the venue survey; until then it lives here.
Distinct from the conference surveys, which cover published/accepted work. Add newest first.
β Back to Home Β· full review catalog: Reviews
In-depth reviews
| Posted | Paper | arXiv | One-line |
|---|---|---|---|
| 2026-08 | [DYNA-2 (World-Action Model, human-video scaling law)](/Heungwoo/research/wiki/Review-Dyna2) | (tech report β self-published, not peer-reviewed) | Dyna Robotics' mixture-of-transformers WAM on ~1M h human egocentric video, no robot data in pre-training; joint next-frame+next-action; published 1kβ1M h power-law fits (RΒ²β0.88β0.93, ~50Γ EgoScale) incl. zero-shot humanβrobot transfer, and 87% vs 46% over its DYNA-1 VLA β company self-reported, unreplicated |
| 2026-08 | [Ο-0 (humanoid loco-manipulation WAM)](/Heungwoo/research/wiki/Review-Omega0) | 2608.06375 | Whole-body humanoid WAM; reconstruction-free latent future prediction (not video) β SONIC control; single model does 11 household tasks at 81.8% vs 44.5% best baseline; ships the 40 h Ο-HOME dataset |
| 2025-11 β 2026-05 | [Stellar VLA (continual skill knowledge)](/Heungwoo/research/wiki/Review-Stellar-VLA) | 2511.18085 | Continual imitation learning for a fixed-size ~1B VLA with 1% replay; Dirichlet-Process self-evolving knowledge space (T-Stellar/TS-Stellar) + knowledge-routed diffusion MoE; SOTA CIL on LIBERO + real dual-arm |
| 2026-04 | [Cortex 2.0 (foresight-planning WAM+VLA)](/Heungwoo/research/wiki/Review-Cortex2) | 2604.20246 | Sereact's modular hybrid: world model generates k candidate futures in visual latent β PRO scores progress/risk/termination β flow action head commits, 30 Hz replan; deployment-reported 0.95β0.98 success / 0 interventions vs Ο0.5 on industrial tasks (>10M-episode fleet) β company-reported |
| 2026-04 | [Being-H0.7 (latent world-action model)](/Heungwoo/research/wiki/Review-Being-H07) | 2605.00078 | BeingBeyond's unified + reactive hybrid: learnable latent queries as reasoning interface; future-informed posterior trains a deployable prior (no visual rollout at inference); InternVL3.5+Qwen3+V-JEPA2.1 on Being-H0.5; LIBERO 99.2%, RoboTwin2.0-Hard 89.6% |
| 2025-12 | [Motus (unified latent-action WAM)](/Heungwoo/research/wiki/Review-MOTUS) | 2512.13030 | THU-ML's unified hybrid: one 8B scheduled Mixture-of-Transformers is a world model / VLA / IDM / video generator on demand; optical-flow latent actions; open weights; RoboTwin 2.0 88.66% vs X-VLA 72.80% / Ο0.5 42.98% |
Where these connect
- Ο-0 extends Humanoid VLA and sits in the World Models WAM taxonomy as the latent-predictive corner β a counterpoint to DreamZero's pixel-video WAM and to MotionWAM/DiT4DiT.
- DYNA-2 is an industry WAM release backed by a self-published technical report (architecture + power-law fits, not peer-reviewed) pushing the EgoScale human-video scaling law and the DreamZero/Ο-0 joint video-action formulation to ~1M-hour, human-only scale β reviewed with an explicit not-peer-reviewed/unreplicated caveat.
- Stellar VLA advances the continual-learning thread from RL for VLA and answers ICML 2026's Pretrained VLAs Resist Forgetting with a structured (not just replay-based) recipe.
β Back to Home