BURGER: Benchmarking and Understanding Robotic Generalization via Embodied Drawing
Some robot models learned to draw unseen combinations of familiar objects as training diversity and demonstration counts grew.
BURGER tests five vision-language-action models using 28,200 simulated demonstrations across 282 drawing tasks. Each task starts with a blank canvas and a language instruction, allowing multiple valid drawings and stroke orders. The experiments reveal distinct failures in language grounding, spatial composition and action generation, with some models benefiting from more training data. A trained policy also ran on a Franka Panda arm, drawing reactively under human intervention.
ArXiv · AI/CL/LG's note
BURGER tests five vision-language-action models using 28,200 simulated demonstrations across 282 drawing tasks. Each task starts with a blank canvas and a language instruction, allowing multiple valid drawings and stroke orders. The experiments reveal distinct failures in language grounding, spatial composition and action generation, with some models benefiting from more training data. A trained policy also ran on a Franka Panda arm, drawing reactively under human intervention.
ArXiv · AI/CL/LG's note
score 4