BURGER turns robotic drawing into a stress test for vision-language-action (VLA) systems: can a robot turn an instruction into a coherent sequence of strokes, then adapt when the canvas changes? The benchmark pairs 28,200 generated demonstrations with simulated and real-arm tests, revealing that current models follow familiar categories and can respond to human edits, but struggle to draw concepts outside their training vocabulary.
What Did the Researchers Build?
BURGER is a benchmark and dataset for testing whether robots can follow language instructions while drawing. Drawing looks simpler than many manipulation jobs, but it still asks a robot to connect words to objects, plan a sequence of movements, and keep track of what it has already put on the canvas. Those demands also appear in tasks such as assembling items or handling objects based on verbal directions.
The researchers built a scalable collection pipeline around sketches from Quick, Draw!, a dataset with 345 object categories. They selected 132 categories considered common in the real world and practical for a robot arm to draw. The benchmark also includes 150 two-object tasks described by spatial relationships: 100 use left-right arrangements, and 50 use vertical relations such as “over” and “under.” With 100 trajectories per task category, the dataset contains 28,200 trajectories in total.
BURGER also includes real-robot data and experiments. Its physical setup uses a Franka Panda arm and two cameras to observe the drawing surface. A small real-robot dataset covers 10 object categories, with 10 demonstrations for each. This makes it possible to test whether training and evaluation methods developed in simulation can support drawing with a physical robot.
What Are the Key Results?
BURGER tests several capabilities separately: following object names, understanding spatial relationships, reproducing different valid drawing styles, scaling with more data, and responding to changes on the canvas. The reported findings show that VLA models can distinguish familiar object instructions in the tested categories. But changing the prompt to name an object outside the training categories does not make the robot draw that unseen concept. In this setup, language following works within learned categories, not as open-ended object knowledge.
Drawing also has multiple valid answers. A bicycle, for example, can be sketched with different shapes or stroke sequences while remaining recognisable. The benchmark includes this variation and shows examples of policies reproducing more than one valid trajectory for a given task, including in real-robot demonstrations.
The physical-robot tests illustrate closed-loop interaction—using new visual observations to decide what to do next. When given a partly drawn cup, the robot adds a missing handle. In another example, a person adds a bicycle’s rear wheel; the robot observes the updated canvas, avoids repeating that part, and continues the drawing. These examples demonstrate useful behaviours, but the reported text does not provide aggregate success rates or numerical comparisons between model families. The evidence is therefore best read as a diagnostic set of findings, not a claim that one model wins by a measured margin.
How Does BURGER Work?
A BURGER task combines a language instruction with the current view of a drawing canvas. During training, each example also includes the robot’s internal position information and an expert action: a command describing how the end effector—the tool tip at the end of the arm—should move. The system learns to connect the prompt and visual state to those movements.
The training trajectories are generated through a drawing pipeline that converts sketch instructions into paths, then replays those paths as robot actions. In evaluation, the policy receives the instruction, canvas image, and robot state, and produces a movement or a short sequence of movements. It must also learn when to stop: an episode ends when the tool is lifted above a set height. That makes termination part of the task, rather than an automatic step after a fixed number of strokes.
The benchmark evaluates more than whether the final image looks roughly like the requested object. Stroke order and geometry matter, and two different paths can both be valid. A general image–text similarity score can miss small but important drawing mistakes, so BURGER also examines the full trajectory and uses a task-specific evaluation metric. This makes it possible to inspect how a policy drew—not just what its finished picture resembles.
For physical tests, two Intel RealSense cameras capture the workspace from different directions. Depth information helps separate the robot arm from the strokes, and observations from both views are combined before being passed to the policy. If a person enters the workspace, the system pauses; after the person changes the drawing and leaves, the robot processes the updated canvas and resumes.
Why Does This Matter for Robotics?
A robot that follows a written instruction must do more than recognise a word. It must ground that word in what it sees, plan an extended action, monitor its progress, and recover when the world changes. BURGER puts those abilities in a setting where the result is easy to inspect and where different valid actions can be compared. That can help engineers identify whether a failure comes from unfamiliar language, a missed spatial relationship, poor movement planning, or a failure to notice a human edit.
The benchmark’s drawing task is not a direct substitute for testing warehouse picking or factory assembly. Its value is as a controlled probe of capabilities that those applications also need. For example, a warehouse robot might need to interpret “place the small bin beside the large one,” then adjust if a worker moves an item. BURGER gives researchers a way to study related language-and-action behaviours without the many physical variables involved in heavier manipulation tasks.
For buyers and operations teams, the results are a reminder that impressive performance on familiar instructions does not guarantee reliable handling of unfamiliar objects or situations. Teams comparing systems should ask how a robot performs on new combinations, whether it observes changes during execution, and how its failures are measured. Those exploring physical platforms can browse used robots for sale while keeping evaluation needs in view.
What Are the Limitations and Open Questions?
BURGER focuses on sketch-level drawing with a tool, not the richer contact and material behaviour involved in painting. Brush pressure, ink flow, and interactions with different surfaces are outside this setup, so results should not be taken as evidence that a robot can handle artistic media more broadly.
The data-generation pipeline also shapes what the models learn. Its rules for choosing stroke order can create consistent trajectories, but those conventions may differ from how people naturally draw. A larger dataset does not automatically remove that bias if the examples share the same assumptions.
Further questions include how well the results carry over to unfamiliar workspaces, different tools, and longer tasks. The described real-robot collection is limited to 10 object categories with 10 demonstrations apiece, and the reported examples do not establish broad performance across those conditions. More varied human demonstrations and quantitative comparisons across task types would help clarify where the approach transfers and where it breaks down.
Frequently Asked Questions
What does BURGER stand for? BURGER is the name of the robotic drawing benchmark; the paper describes it as a benchmark for understanding robot generalization through embodied drawing.
What does the robot receive as input? It receives a language instruction, a visual observation of the canvas, and its robot state. It then predicts movements for the drawing tool.
Can BURGER robots draw objects they have never seen in training? The reported tests found that models struggled to produce requested object categories outside their training set. Familiar categories were handled more reliably.
Can the robot respond when a person changes the drawing? Yes. In demonstrations, the robot pauses while a person edits the canvas, then uses the updated visual observation to continue the task.
What Does BURGER Show Overall?
BURGER makes language-conditioned drawing a practical way to inspect how VLA systems connect instructions, visual feedback, and long sequences of movement. Its real-robot examples show useful adaptation to human edits, while its unseen-category tests expose a clear boundary: following familiar labels is not the same as generalising to new concepts.