Physical AI · evidence and evaluation
Domain randomization: what it varies, and what transfer it buys
Domain randomization trains a model across randomly varied simulated conditions so that reality looks like one more variation. It buys robustness to the parameters that were randomized, and nothing about the ones that were not.
How it works
Tobin and colleagues randomized a specific list during training: the number and shape of distractor objects, the positions and textures of objects, the textures of table, floor, skybox and robot, the camera’s position, orientation and field of view, the number and properties of lights, and image noise. That list is the claim. Anything held fixed in simulation is a property the model may silently depend on — in their setup, table height was fixed, and the estimate was a planar one from an uncalibrated monocular camera. When someone says a policy was trained with domain randomization, the informative question is always which parameters, over what ranges.
A concrete case
A grasping policy randomized over textures and lighting still fails on a transparent object, because the rendering never produced refraction. The randomization list did not include the physics that the real object exhibits.
What this establishes
That the technique and its randomized parameter list are published, and that transfer was demonstrated for the paper’s own task.
What it does not
One task, one setup, one set of ranges. It is not evidence that randomization transfers a different task, and simulated success remains simulated.
Questions worth asking
- Ask for the randomized parameter list and its ranges, not the phrase.
- Ask which physical phenomena the simulator does not model at all.
- Ask whether real-world evaluation used objects and lighting outside the randomized ranges.
Sources
- Tobin et al. — Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World ↗
The method randomises stated aspects of the simulated scene during training — number and shape of distractor objects, object positions and textures, table, floor, skybox and robot textures, camera position, orientation and field of view, lighting, and image noise — so that the real world appears to the model as one more variation.
Boundary: The paper demonstrates transfer for its own task and setup, with a fixed table height and an uncalibrated monocular camera. Randomising a list of parameters is not evidence that any other task transfers.
Locator, anchor and reuse basis
Read at: §III-A Domain randomization. Inspected 2026-09-19. Original paraphrase and link to the open preprint.
Verify by searching the source for:
position, orientation, and field of view of the camera
. If that phrase is not there, or does not carry the meaning stated above, this citation is wrong and we want to know. - Ha and Schmidhuber — World Models ↗
A controller can be trained inside a learned generative model of an environment and then transferred back; the authors discuss how an agent can exploit imperfections of that learned model, and use a temperature parameter to make the imagined environment harder to exploit.
Boundary: The environments are video games. Nothing in the paper concerns physical contact, hardware safety, or a robot acting among people.
Locator, anchor and reuse basis
Read at: §4.4 Transfer policy to actual environment; surrounding discussion of model imperfections. Inspected 2026-09-19. Original paraphrase and link to the open preprint.
Verify by searching the source for:
the tradeoff between realism and exploitability
. If that phrase is not there, or does not carry the meaning stated above, this citation is wrong and we want to know.
Continue
Elsewhere on this site
Maha Strategies publishes explanation and evaluation method. We build no robots, run no physical experiments, and report no benchmark results of our own. Hardware, safety and evidence-intake questions live in the robotics section.