A new way to test how robots learn to draw from language
The BURGER benchmark uses a scalable pipeline to test robotic drawing tasks in simulation and the real world.
Researchers introduced BURGER, a language-conditioned robotic drawing benchmark. It uses a prompt-driven pipeline to create continuous drawing trajectories in simulation and the real world.
A robot arm starts from a blank canvas. It draws objects or spatial compositions specified by language. Each instruction allows many valid drawings and stroke orders.
The study used 28,200 simulated demonstrations spanning 282 drawing tasks. This was used to evaluate five representative vision-language-action (VLA) models.
The benchmark allows analysis of language grounding, spatial composition, and action generation. These tests reveal distinct failure patterns across VLA models.
Experiments suggest some VLA models benefit from greater task diversity and demonstration scale. They may capture multimodal drawing behaviors and generalize to unseen compositions of familiar objects.
A trained policy was deployed on a Franka Panda arm. The arm demonstrated the ability to draw reactively under human intervention.
Key facts
- BURGER uses 28,200 simulated demonstrations across 282 drawing tasks.
- The benchmark tests five representative vision-language-action (VLA) models.
- A trained policy was deployed on a Franka Panda arm to draw under human intervention.