World Model & VLA Breakthroughs: Top Papers from NVIDIA, ByteDance, Tsinghua and Others

DreamDojo: Generalist Robot World Model Trained on Large‑Scale Human Videos

Trained on 44,000 hours of first‑person video (DreamDojo‑HV dataset), the model incorporates latent actions to address the scarcity of action labels. It enables real‑time, physics‑aware robot simulation suitable for open‑world tele‑operation and planning. The dataset is the largest human‑interaction video collection used for world‑model pre‑training.

LingBot‑World: Open‑Source Video‑Generated World Simulator

Built on video‑generation techniques, LingBot‑World maintains high‑fidelity dynamics across diverse scene types (realistic, scientific, cartoon). It provides multi‑minute prediction horizons with long‑term memory and supports real‑time interaction at 16 fps with sub‑second latency.

Agent World Model (AWM): Code‑Driven Synthetic Environments for Agentic RL

AWM generates synthetic environments via code, offering 1,000 diverse scenes and 35 toolkits. It outperforms LLM‑based simulators and improves out‑of‑distribution generalization through executable, database‑backed state representations. The training dataset expands from 100 seed domains (popular websites) to 1,000 CRUD‑oriented scenes, each representing a real‑world application (e‑commerce, CRM, banking, tourism) and filtered to ensure diversity.

BagelVLA: Interleaved Vision‑Language‑Action Generation for Long‑Horizon Manipulation

BagelVLA unifies vision, language, and action generation. Residual Flow Guidance fuses language planning with visual prediction, yielding precise, low‑latency action generation. Experiments on complex multi‑stage manipulation tasks show significant performance gains over baseline VLA models.

ACoT‑VLA: Action Chain‑of‑Thought for Vision‑Language‑Action Models

ACoT‑VLA introduces an Action Chain‑of‑Thought reasoning layer that jointly leverages coarse‑grained intent and latent action priors. This design achieves state‑of‑the‑art results on the LIBERO, LIBERO‑Plus, and VLABench benchmarks, surpassing prior VLA approaches.

World‑VLA‑Loop: Closed‑Loop Learning of Video World Model and VLA Policy

World‑VLA‑Loop iteratively refines a video world model and a VLA policy using failure feedback. Leveraging the SANS dataset (compiled from ManiSkill, LIBERO, and real‑robot recordings), the framework improves action‑following accuracy and raises real‑robot task success rates by 36.7 % in simulation experiments.

The six works illustrate a shift in embodied AI toward agents that learn and decide within generative, interactive worlds, using large‑scale video data, synthetic environment generators, and closed‑loop refinement mechanisms.