Data and benchmarks
Can camera processing change a robot's apparent manipulation skill?
A robot policy acts on processed camera images, not directly on a scene. RawVLA tests how that processing affects actions and reported task success, especially when lighting changes.
In RawVLA's controlled tests, the camera pipeline changed what a robot policy saw and how often it completed manipulation tasks under altered light. This is one author-reported study, not evidence that all robots will benefit from the same camera software.
The image is part of the test
A vision-language-action policy receives an RGB image produced by camera processing. Gain, sensor noise, color response, tone mapping and bit depth can all affect that conversion. The processing choices can change the image the policy uses for its next action.
The RawVLA preprint treats this conversion as a variable rather than a neutral step. In one sensitivity test, the researchers changed one processing dimension at a time while keeping the scene, task and robot state fixed. They report that processing choices affected action predictions and manipulation success. That comparison matters because a change in score need not come from a change in the robot's learned policy or the task itself.
RawVLA is their proposed adaptive converter. It renders RAW observations into images for a frozen downstream policy, concentrating on imaging factors useful for action. The claim is therefore about the camera-to-policy interface in the tested setup, not a robot learning a new skill from scratch each time the light changes.
What the benchmark and physical trials measured
RawVLA-Bench varies illumination across five regimes, from extremely low light through normal light to extreme overexposure, in LIBERO and RoboTwin 2.0. The simulated evaluation uses task success as its outcome and the authors report 50 rollouts per task with different seeds and starting positions. The design lets them compare image processing under controlled visual changes, but simulated RAW images are approximations of sensor data.
In a small physical comparison, the authors report that RawVLA performed much better under low light than the default image processor, while performing similarly under normal light. The comparison covers their tested indoor manipulation tasks and lighting conditions. It is not a general reliability figure for deployed robots. Readers need the imaging conditions alongside task success to understand what was tested.
What this does not establish
The simulated experiments use three-channel pseudo-RAW observations rather than camera-native mosaiced measurements. They do not capture the full range of sensor responses, demosaicing artifacts, lens effects or proprietary camera pipelines. The authors also trained a separate adaptive processor for each downstream policy. Their physical evaluation covers a small set of indoor tasks and lighting conditions.
For a claim that a robot has learned a skill, ask what the camera delivered to the policy as well as how the task was scored. RawVLA gives one concrete reason: changing the image pipeline can change measured behavior even when the policy and task stay the same.