ABot-World-0 rolls out 720p interactive worlds on a single RTX 5090

Fan Jiang · hf · 2026-07-22

ABot-World-0 streams interactive world rollouts at 720p on a single RTX 5090

The paper introduces ABot-World-0, an action-conditioned video world model built for real-time, long-horizon closed-loop interaction.

Training data comes from a multi-source pipeline spanning AAA games, simulation engines, and internet videos.
WorldExplorer uses agent-driven collection guided by training feedback, while the data pipeline applies 14 deterministic quality checks, VLM-based assessment, and synchronized action/text annotation.
The model is distilled from a bidirectional teacher into a causal student using teacher forcing and ODE distillation.
A method called LongForcing aligns long student self-rollouts with an extended-horizon teacher to reduce distribution shift and autoregressive drift.
The system uses raw keyboard actions for both roaming and third-person interaction, plus reference-character memory for persistent appearance consistency.
For deployment, the authors co-design a streaming stack with a lightweight VAE decoder, efficient attention, memory-aware scheduling, and low-bit DiT inference.
On optimized low-bit settings, it streams 720p video at up to 16 FPS on a single NVIDIA RTX 5090, with 1.2 s action-to-first-frame latency and about 19 GiB peak VRAM.
Results on WorldRoamBench and long rollouts show competitive controllability and coherent long-horizon evolution.

More from Embodied