TriWorldBench: The First Three‑View Embodied World‑Model Benchmark

When robots use multiple cameras, inconsistencies between views can cause a model to generate videos that do not represent the same underlying world. This raises the fundamental question: how can we tell whether a world model truly understands the robot’s environment?

In response, Peking University, Tsinghua University, Beihang University, Shanghai Jiao‑Tong University, and the University of Science and Technology of China jointly launched the TriWorldBench Challenge , the first benchmark that evaluates embodied world models from three synchronized perspectives—head view, left‑wrist view, and right‑wrist view.

From Video Generation to World Understanding

Traditional video‑generation metrics focus on visual clarity, fidelity to description, and smoothness. For embodied models, these criteria are insufficient because robots operate in a dynamic, physics‑driven world observed by multiple cameras that together form a complete perception.

A reliable embodied world model must not only produce plausible videos for each view but also ensure that all cameras depict the same world state.

TriWorldBench Evaluation Dimensions

01 Multi‑View Consistency : The three videos must align on arm motion, grasping and contact phases, and object category, appearance, and state. The model must describe a single coherent robot operation rather than three independent plausible scenes.

02 Task Execution Ability : Beyond visual quality, the model is judged on whether the robot actually completes the instructed task—correct command execution, expected arm motion, and reasonable grasp‑move‑release actions. Head view assesses overall task progress, while wrist views verify fine‑grained contact stability.

03 Physical World Understanding : The model must capture 3‑D spatial relationships, motion dynamics, and cross‑view spatial correspondence.

Benchmark Structure and Scoring

The benchmark contains 500 synchronized three‑view episodes covering 50 robot manipulation tasks. It evaluates six dimensions—multi‑view consistency, task alignment, physical/3‑D consistency, motion quality, temporal consistency, and visual quality—producing 19 signals that are aggregated into a single TWB‑Score.

Each view is first compared to its ground‑truth camera, then semantic head‑wrist reasoning and joint three‑view Q&A verify arm state, action stage, contact relations, and object compatibility. Scores are assigned to the most reliable camera for each metric, preventing a high‑quality head view from masking wrist‑view failures.

The benchmark also incorporates a STATE annotation that marks which arm should be moving or stationary at each moment; unnecessary motion or frozen wrists incur penalties, ensuring that motion richness reflects true task understanding.

Diagnostic Value

TriWorldBench does not simply average the three videos into a single number. Instead, it provides a full evaluation chain—from synchronized data to multi‑level results—allowing researchers to pinpoint whether low scores stem from task execution, motion errors, visual quality, or cross‑view inconsistencies.

In visual‑language model (VLM) semantic tests, human expert scores are used to align the evaluation protocol before freezing the rules for official testing.

Open Participation

The challenge is open to global research teams, including embodied world‑model groups, embodied intelligence labs, video‑generation teams, and multi‑view generation & understanding researchers. Participants can submit models to the public leaderboard, receive detailed diagnostic feedback, and compare approaches on a common, scientifically rigorous platform.

TriWorldBench aims to move the field from “looks like” video generation toward truly usable, world‑understanding models that can predict and act in real environments.