ACT

Open weightsStanford (ALOHA) · baseline, not a foundation model

3 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)

   48.3%

   Real robotGrade B · Paper or report with method (rubric or error bars)200 trialsALOHA · 4 tasks

   Clean dish 86%, sponge 85%, unzip 22%, fold 0%. On the SO-101 benchmark it averaged 33.75%.

   Source: Benchmarking VLAs on ALOHA, arXiv 2511.11298 · Table 3, macro-average

2. Robustness

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

   5.5%

   Real robotGrade B · Paper or report with method (rubric or error bars)200 trialsALOHA · unseen objects

   Collapses off its training setup: 42.8 points below in-distribution.

   Source: Benchmarking VLAs on ALOHA, arXiv 2511.11298 · Table 5 (mean of 4 tasks)

3. Instruction

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

   Owed

   Not reported.

4. Recovery

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

   6.5%

   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

   Not reported.

6. Interventions

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

   Owed

   Not reported.