AI

Rhoda Scales Web Video and the Bearings Come Out Faster

Shar Hendrix 4 min read

Rhoda AI asked a blunt question: if you pre-train a video model on ordinary web clips, with no robot actions in them, does a bigger or longer run actually help a factory cell? On 10 September, Tongzhou Mu and the Rhoda team published the measurement. Humanoids Daily picked it up on 15 September.

The task is not a cube on a table. A customer unpacks more than 1,000 bearing boxes a day by hand. Each box weighs 10 kg. The robot has to lift it, find the tab, open the lid, dump the bearings, and sort cardboard, plastic, and paper. A trial counts only if every goal is met with no human help and the cycle finishes under 100 seconds.

Two gray robot arms with blue rings unpacking bearings from gray totes on a conveyor
Rhoda’s bearing-unpacking cell, shown at 1× speed. Source: Humanoids Daily / Rhoda AI.

Size first, then compute

Rhoda’s Direct Video-Action model predicts future frames, then a separate inverse-dynamics model turns those frames into joint commands. That inverse model stays fixed. Only the pre-trained video net changes.

Four sizes, one recipe, full demonstration set:

SizeAt-speed completionTrials
XS3.7%1/27
S65.0%67/103
M75.3%195/259
L84.7%94/111

The team stopped XS after 27 trials once it was clearly worse. They say the L model has not saturated the metric.

Holding size at M and scaling pre-training compute from 0.08× to of the full run moved the rate 57.8% → 67.3% → 74.5% → 75.3%. The last step is flat at full task data. Cut the demonstrations and the gap opens: at 25% of the post-training set, the spread between the largest and smallest compute budgets reaches 31 points.

Scatter plot showing higher pre-training quality aligned with higher robot task performance from XS to L
Held-out web-video quality versus on-robot at-speed completion. Source: Rhoda AI research post.

A cheap pre-training score, DINO FD on held-out web video, ranked the seven checkpoints in the same order the robot did. Lower distance, higher completion, from 4% on the left to 85% on the right.

The evaluations ate more than 200 robot hours. Each policy got about 100 to a few hundred trials under a written SOP. The 100-second cutoff is stricter than the customer’s own limit, on purpose, so the score does not saturate.

Rhoda lists the limits in the same post: one task, one cell, one post-training run per condition. Larger models also saw more video, so the size sweep is not compute-matched.

A Human’s Take

This is the rare scaling chart I trust because they published the failure modes. Bigger models mostly win on time, not on “did the bearings land in the tote.” They retry less and move faster. Waste mis-sorts even tick up a bit when the policy commits before it has looked long enough. That is a real factory trade, and they measured it instead of waving a success rate.

Sources