Updated from microAGI report supplied by user; worker to verify published figure table; MicroAnnotationBench excluded until separately checked.

Date: 2026-10-20
Type: ARTICLE
Topics: Benchmarks, Egocentric Video, Robot Planning, Vision Language Models
Deck: GPT-6 Astra scored 65.6% on a 200-clip test of immediate next-action prediction, with hand assignment and object selection included in grading.
URL: https://korthosrobotics.com/research/microagi-microplanbench-next-action-video-benchmark-2026
Referenced profiles: microagi (https://korthosrobotics.com/ecosystem/company/microagi)

Image: microAGI
microAGI has published MicroPlanBench, a benchmark of how well multimodal models predict a person’s next action from first-person video. The report, dated 14 October, evaluates 13 models on 200 clips covering 10 work domains in Europe and North America. GPT-6 Astra led the company’s results at a 65.6% mean pass rate, closely followed by GPT-6.1 Sol at 65.5%.
Each example ends just before an atomic action, such as moving an object to a destination with a particular hand. Models receive eight seconds of 1280×720 video sampled at ten frames per second and must describe the immediate next action in a sentence. Every case is run five times to account for variation between predictions. The underlying clips total just under 27 minutes.
The evaluation set was filtered to moments where reviewers agreed that one action would reasonably advance the task. A vision-language model first screened candidate clips and generated four alternative actions. Human reviewers then assessed the five options without knowing which action actually followed. Selected examples received corrected labels for manipulation, objects, destination and left- or right-hand use.
A GPT-6 Astra judge checks predictions against those labels. Correct answers must match the action, objects, destination and hand assignment and describe only the next step. microAGI tuned its judging prompt against 1,000 human-labelled predictions and reports 99.2% consistency between repeated gradings across six judging passes.
The leading model failed on 344 of 1,000 predictions. The report groups failures by incorrect manipulation, hand assignment, and object or destination, with manipulation accounting for the largest share. Its examples span packing, food preparation, retail, motor repair and tailoring. The evaluation measures task-step selection from observed human activity.