Summary
Video world models are emerging as scalable alternatives for evaluating generalist robot policies without physical hardware. RoboWorld is an automated evaluation pipeline pairing a fast autoregressive video world model with a task-progress-aware vision-language model scorer, enabling reliable policy benchmarking at scale without robot access.
Key Contributions
- Automated evaluation pipeline: autoregressive video world model + VLM-based task-progress scoring
- Fast inference enabling scalable policy evaluation across diverse tasks
- Task-progress-aware scoring that goes beyond binary success/failure
- Reduces evaluation cost and engineering burden of physical robot benchmarking
Significance
RoboWorld validates world models as policy evaluation tools—a complementary use case beyond planning and data generation—potentially enabling continuous policy monitoring and large-scale benchmarking that would be prohibitive with physical robots.