Astronex-World 1.0: Real-Time Interactive World Model Foundation
cs.CV · cs.AI · cs.RO
Quick verdict
5B open world model trained on 2x L20 GPUs streams 24fps and scores 70.0 WBench Full, near 22B rivals; detailed recipe, but self-run evals temper comparisons.
Summary
This technical report presents Astronex-World 1.0, a 5B-parameter controllable video world-model family built on the Wan2.2-TI2V-5B prior and released in bidirectional and causal forms. A 64-dimensional per-frame action stream with 32 embodiment IDs modulates every Transformer layer, PRoPE injects camera intrinsics and extrinsics into all 30 layers, and timed text events can be inserted into an ongoing rollout. A five-stage recipe — bidirectional control adaptation, block-causal teacher forcing, UniPC trajectory distillation, mixed-domain causal SFT, and asymmetric DMD/DMD2 with a motion-preservation term — reportedly runs entirely on two NVIDIA L20 48GB GPUs. The causal model streams 832×480 video at 24 fps on a single L20 via block-causal attention, a 20-frame KV window with four persistent sink frames, and eight-step UniPC sampling. Self-evaluated WBench scores are 73.5 (Navi 158) and 70.0 (Full 289) using a doubao VLM judge, above the 13.6B LongCat-Video and 14B Helios, 0.9 below the 22B LTX-2.3, and above the same-prior YUME 1.5. Acknowledged limitations: weak event and subject-action interaction (Interaction 47.6 on Full), long-horizon color and structural drift, and severe DMD motion suppression (an internal diagnostic shows motion amplitude falls to 19% of the bidirectional teacher). VBench covers only 240/6,220 T2V and 118/5,590 I2V videos, and no variance estimates or ablations are reported.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A 5B causal video world model streams 832×480 video at 24 fps in real time on a single NVIDIA L20 using block-causal attention, cross-block KV caching, a 20-frame window with four persistent sink frames, and eight-step UniPC sampling (Sec 7.1).
- The entire five-stage post-training recipe — control adaptation, block-causal conversion, trajectory distillation, mixed-domain SFT, asymmetric DMD/DMD2 — runs on two NVIDIA L20 48GB GPUs (Sec 6.2, Table 4).
- The causal release scores 73.5 on WBench Navi 158 and 70.0 on Full 289, above 13.6B LongCat-Video (69.9), 14B Helios (69.7), and same-prior YUME 1.5 (68.9), and 0.9 below 22B LTX-2.3 (70.9) (Sec 8.1-8.2).
- A unified control interface (PRoPE camera geometry in all 30 layers, 64-D action stream with 32 embodiment IDs, timed text events) operates on one shared trunk, with event edits staying local to the rollout (Sec 5.3-5.4).
- The reserved 64-D action input and [F,64] action-sequence output head make the foundation post-trainable for embodied intelligence and autonomous driving (Sec 8.7).
Strengths
Unusually complete training disclosure
Sections 6.2-6.7 and Table 4 report per-stage data, initialization, LoRA rank 64/alpha 128, learning rates (e.g., Stage IV trunk 2e-6 vs action 2e-5), update counts, and effective batch sizes — enough to audit the 2×L20 efficiency claim, which most tech reports never provide.
Honest failure and incompleteness reporting
Section 8.8 quantifies DMD motion suppression on an internal sequence (88.8 px bidirectional → 16.6 px at DMD1545, 19%), Section 7.3 states VBench is only ~4%/~2% complete, Figure 10 shows named failure cases, and Table 6 visibly keeps the model below HY-Video 1.5 (74.3 vs 70.0).
Coherent causal-conversion pipeline
Event-insertion locality is mechanism-tested
Weaknesses
Cross-model comparison is not like-for-like
VBench evidence is thin and possibly unrepresentative
No uncertainty and no ablations
Implausibly small or incompletely accounted training budgets
Real-time claim is narrow
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
Astronex-World 1.0 (arXiv:2609.20034) is a technical report from Astronex Robotics describing a 5B-parameter controllable video world-model family built on Wan2.2-TI2V-5B, released in bidirectional (full attention, checkpoint restore-2200) and causal (block-causal, checkpoint DMD1545) forms. The control interface comprises PRoPE camera intrinsics/extrinsics injected in all 30 layers, a 64-dimensional per-frame action stream with 32 embodiment IDs modulating every layer (Table 1), timed text events inserted at a chosen latent frame (Sec 5.4), and a reserved [F,64] action-sequence output head. Central claims: (1) the full five-stage recipe runs on 2× NVIDIA L20 48GB (Sec 6.2); (2) the causal model streams 832×480 at 24 fps on one L20 (Sec 7.1); (3) WBench Navi 158 = 73.5 and Full 289 = 70.0, above 13.6B LongCat-Video (69.9) and 14B Helios (69.7), within 0.9 of 22B LTX-2.3 (70.9), and above same-prior YUME 1.5 (68.9) (Tables 5-6); (4) the interfaces make the foundation post-trainable for robotics and driving (Sec 8.7).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Astronex-World 1.0?
An open, 5-billion-parameter controllable video world model built on the Wan2.2-TI2V-5B prior, released in two forms: a bidirectional model for full-context generation and a causal model that streams video interactively in real time. It accepts a text prompt or an initial image and follows camera trajectories, a 64-dimensional action stream with an embodiment identifier, and timed text events inserted mid-rollout.
How does Astronex-World 1.0 compare to larger world models?
The authors self-report WBench Full 70.0 (and Navi 73.5), higher than 13.6B LongCat-Video (69.9) and 14B Helios (69.7), 0.9 points below 22B LTX-2.3 (70.9), and above the same-prior 5B YUME 1.5 (68.9). Caveat: peer numbers come from a leaderboard snapshot while Astronex-World was evaluated by the authors themselves with a doubao VLM judge, so the comparison has not been independently calibrated.
What hardware does Astronex-World need to run in real time?
The causal model streams 832×480 at 24 fps on a single NVIDIA L20 48GB data-center GPU using eight-step UniPC sampling, eight-latent-frame blocks, and cross-block KV caching; training used two such GPUs. The paper reports no latency figures for consumer GPUs.
Can Astronex-World control robots or drive cars?
Not out of the box. The 64-dimensional action input and a per-frame action-sequence output head are reserved interfaces: the paper demonstrates no robot or driving post-training, and lists connecting them via domain post-training as future work.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-13 are available only as captions in the extracted text; qualitative claims (identity preservation, drift cases, consistency crops) could not be visually verified.
- Section 3 (System Overview) is interleaved with figure-extracted labels, so some pipeline prose may be missing or garbled in the extraction.
- Leaderboard scores for peer models (Table 6) cannot be verified from the provided text and are taken at face value; only internal arithmetic was checked.
- This is a non-peer-reviewed vendor technical report; affiliations, 2026-dated references, and website/repo claims were accepted as stated.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.