SmolVLA

Open weightsHugging Face

2 of 6 axes published · values checked against their primary sources on 23 Oct 2026

1. Success

   Does it do the trained task? (% of trials, trained setup)

   32.5%

   Real robotGrade B · Paper or report with method (rubric or error bars)80 trialsSO-101 · 4 tasks

   Independent benchmark. SmolVLA's own paper reports 78.3% on its own SO-100 tasks (vs 61.7% for π0).

   Source: SO-101 VLA benchmark, arXiv 2606.08881 · Results table, average

2. Robustness

   Does it survive the cell drifting? (% under the perturbation set)

   Owed

   Not reported. SmolVLA is not in LIBERO-Plus or IndustrialVLA-Bench.

   Checked: IndustrialVLA-Bench, arXiv 2609.25562

3. Instruction

   Does rewording the instruction break it? (% under paraphrase)

   Owed

   Not reported.

   Checked: SmolVLA paper, arXiv 2506.01844

4. Recovery

   Does it fix its own mistakes? (% of injected failures recovered)

   3.2%

   Real robotGrade B · Paper or report with method (rubric or error bars)SO-101

   Share of failures recovered without help.

   Source: SO-101 VLA benchmark, arXiv 2606.08881 · Recovery results

5. Speed

   Is it as fast as a trained person? (human cycle time ÷ robot cycle time)

   Owed

   No human comparison. Its own paper: asynchronous inference finishes tasks 30% faster than synchronous (9.7 s vs 13.75 s) — against itself, not a person.

   Checked: SmolVLA paper, arXiv 2506.01844

6. Interventions

   How often does a person step in? (per 1,000 cycles)

   Owed

   Not reported.

   Checked: SmolVLA paper, arXiv 2506.01844