Apptronik's Apollo made seven straight throws from a simulation Gemini 4 Argon built out of video

The weeks-to-hours figure is Google's, taken on what the company calls an internal robotics workflow ^[11]. For it to transfer to another team, that team's task would have to look like this one: a throwing task with one participant, clear enough in video for a model to read and simple enough for a physics simulator to represent. The baseline would have to match too. Your manual setup would need to take weeks for the same reasons Google's did.

The write-up lists the dependencies itself: quality of the input, accuracy of the simulation, the target hardware, and testing in real conditions ^[7].

The calibration step is where we would look first. The simulator holds the environment and the participant's kinematics ^[2], and every training episode inherits whatever error sits in that model. Domain randomization varies the simulated conditions so training does not depend on one fixed scene ^[5]. In our view, that can absorb small calibration errors. A simulation that misrepresents the scene itself would still train the wrong behavior.

The 4,500-episode figure needs care. The dev.to write-up attributes it to Google's related developer material, not to the announcement ^[4]. It does not connect that count to the Apollo trial, so a reader cannot tell how many episodes sat behind the seven throws.

The hardware result is one trial ^[3]. A streak of seven is compatible with any number of earlier misses. Turning it into a success rate would take the attempt count, the throw distance and the scoring rule.

The write-up draws the same boundary. The verified material covers an internal demonstration, and performance across other environments, robot types, objects or safety-critical tasks is outside what it shows ^[6]. The primary source is Google's Gemini 4 Argon announcement, which the write-up summarizes ^[9].