Model comparisons

Two models, one decision

Each page puts two policies against each other on the things that decide a project: what card the run needs, how long the control loop is, and how much data it takes before the arm does the task.

  • 6 comparisons
  • Numbers from the training backend
  • One dataset format for all five

GR00T N1.7 vs Pi0.5: speed against contact precision

GR00T N1.7 or Pi0.5 for an SO-100: 152 ms against 485 ms per action step, batch 32 against batch 1, seed or no seed, and which one your task needs.

GR00T N1.7: 80 GB card, 152 ms · Pi0.5: 80 GB card, 485 ms

GR00T N1.7 vs SmolVLA: an 80 GB card against a 24 GB one

GR00T N1.7 needs an 80 GB card at 4 to 12 USD a run, SmolVLA a 24 GB card at 1 to 3. Which one your dataset size justifies, and when the upgrade pays.

GR00T N1.7: 80 GB card, 152 ms · SmolVLA: 24 GB card, 245 ms

ACT vs SmolVLA: specialist against generalist on the same card

ACT and SmolVLA both train on 24 GB for 1 to 3 USD. One is 80 million parameters from scratch, the other 450 million pretrained with language.

ACT: 24 GB card, 20 ms · SmolVLA: 24 GB card, 245 ms

Pi0.5 vs SmolVLA: when flow-matching precision pays for the extra hours

Pi0.5 costs four times more per run and twice the latency of SmolVLA. The failure pattern that justifies the upgrade, and the ones that do not.

Pi0.5: 80 GB card, 485 ms · SmolVLA: 24 GB card, 245 ms

ACT vs GR00T N1.7: the two extremes of the catalog

ACT is 80 million parameters from scratch at 20 ms. GR00T N1.7 is 3 billion pretrained at 152 ms on an 80 GB card. One question decides which you need.

ACT: 24 GB card, 20 ms · GR00T N1.7: 80 GB card, 152 ms

GR00T N1.7 vs GR00T N1.5: which version to select

N1.7 supersedes N1.5 on every axis that matters. The only reason to select N1.5, and the defaults trap that makes the gap look larger than it is.

GR00T N1.7: 80 GB card, 152 ms · GR00T N1.5: 80 GB card, 165 ms

All five side by side

The same numbers the training backend uses. Every model reads LeRobot v2.1 datasets, so switching models does not mean recording again.

ModelGPUInferenceDefault stepsTypical run
GR00T N1.7NVIDIAA100 80 GB or H100 80 GBabout 152 ms20,000about 4 to 12 USD
GR00T N1.5NVIDIAA100 80 GB or H100 80 GBabout 165 ms2,000about 4 to 12 USD
Pi0.5Physical IntelligenceA100 80 GB or H100 80 GBabout 485 ms30,000about 4 to 12 USD
SmolVLAHugging FaceRTX 4090 or any card with 24 GBabout 245 ms20,000about 1 to 3 USD
ACTStanford (ALOHA)RTX 4090 or any card with 24 GBabout 20 ms100,000about 1 to 3 USD

Prices are estimates: the pool rents GPUs on a spot market and selects them by VRAM, so the hourly rate moves day to day. Details per model under /policies, the runnable instructions under /train.

The arm decides less than you think

A real SO-100 is online and free to drive in the browser. The model you pick matters more than the hardware under it.