RVAB Β· Robot Video Annotation Bench

Frontier-quality video annotation
at a fraction of the cost

We benchmarked 8 vision-language models across 4 robot & human-interaction tasks, then built an inference method that routes work intelligently β€” matching frontier accuracy while spending a third of the budget. The same pipeline powers a second-level, audited annotation deliverable.

$3.04
per video-hour, measured (G3.1-Pro dense)
72.7%
accuracy @ 35% of frontier cost
1,552
labeled segments across 20 clips, second-level
16
structured fields per segment, EN + δΈ­ζ–‡
The headline result

Cheaper and better β€” not a trade-off

On the Cosmos RoboFail benchmark, our agreement-gated cascade (M13) lets free open-source models handle the unambiguous clips and only escalates the hard ones to a frontier model. The result sits in the top-left "win" zone: frontier-level accuracy at ~β…“ the API spend.

64% 66% 68% 70% 72% 74% 76% 0Γ— 0.5Γ— 1.0Γ— 1.5Γ— 2.0Γ— frontier API calls per clip β†’ cost accuracy β–² cheaper + better Qwen-2.5-VL-7B (open, free) 65.7% @ 0.0Γ— Gemini 3.1 Pro (solo) 68.7% @ 1.0Γ— β˜… Ours: M13 cascade 72.7% @ 0.35Γ— β˜… Ours: M13 + ensemble 73.7% @ 0.71Γ— Full ensemble (max acc) 75.8% @ 2.0Γ—
open-source (free) frontier solo our method max-accuracy ensemble
Why it works: when cheap models agree, the clip is usually visually unambiguous β€” and their consensus (70.3%) actually beats a frontier model solo (64.1%) on exactly those clips. We only pay for frontier inference on the 35% of clips where models disagree. Statistically validated with paired-bootstrap CIs.
The benchmark

8 models Β· 4 tasks Β· 36 findings

No single model wins everywhere β€” evaluation is task-conditioned. Frontier models lead on image MCQ and dense temporal grounding; open-source models surprisingly lead on free-form RoboVQA. Our method picks the right tool per task.

ModelEPIC dense
(IoU / label)
RoboVQA Cosmos
RoboFail
Robo2VLM
Gemini 3.1 Pro52.4 / 48.941.169.065.8
Gemini 3-Flash50.8 / 38.739.867.067.1
Gemini 2.5 Pro55.3 / 35.238.961.0β€”
Cosmos-Reason1-7B11.1 / 10.955.661.041.7
Qwen-2.5-VL-7B16.0 / 13.450.765.046.8
InternVL3-8B21.8 / 17.947.363.645.5

Bold-cell winners vary by column. Full numbers, significance markers and methodology in the technical report. EPIC n=15 (segment-level n=140), RoboVQA n=63, Cosmos n=100, Robo2VLM n=345.

The deliverable

Second-level dense annotation, audited

Every long clip is segmented at second-level granularity with a 16-field bilingual schema β€” motion phase, contact state, object state-change, sub-goal, spatial relations and a 2–4 sentence description per segment. A second model pass scores confidence and flags only disagreements for human review.

$3.04/hr
measured cost per video-hour (total $2.20 for 43.4 min)
4.9s
mean segment length (second-level)
25.4
words of description per segment (avg)
16 fields
structured labels per segment, EN + δΈ­ζ–‡
00:09.5 – 00:11.5
insertin-contact
insert strap into buckle
The right hand aligns the buckle's slot with the strap end and pushes it through; the left hand keeps the strap taut. state: strap separate β†’ partially threaded
00:11.5 – 00:14.0
manipulatein-contact
pull strap taut through buckle
Both hands draw the strap through until the buckle seats against the loop. sub-goal: fasten the buckle
β–Ά Scrub the live annotations
Pilot Β· external 15-clip set (试采15村)

Same pipeline, fresh headset capture

15 egocentric clips (~40 min total) from an Apple-Vision-Pro-style headset rig β€” stereo HEVC RGB + 26-joint dual-hand tracking + head pose + IMU. Annotated end-to-end with Gemini 3.1 Pro at 100 % timeline coverage, using the same dense-temporal prompt + sanitiser as the main deliverable, with empirical hand-active intervals lifted straight from the headset's own CSV feed as a sensor-side sanity check.

15 / 15
clips at 100 % timeline coverage
815 segments
dense, bilingual, 16-field schema
$5.69
total API spend ($9.35 / video-hour)
3 subjects
kitchen / office / tea-ceremony activities
β–Ά Open the Somantis pilot studio
Pilot Β· sample01 (3 GoPro clips, ~30 min)

Same pipeline, raw GoPro footage

3 GoPro 1080p/60fps videos (~30 minutes total) covering hand-washing, cardboard-box assembly, and ventilation-fan assembly. End-to-end on the video-mode Gemini 3.1 Pro path with the same dense-temporal prompt β€” 100% timeline coverage, bilingual EN+ZH 16-field schema, ~$3.32 / video-hour measured.

3 / 3
clips at 100% timeline coverage
373 segments
dense, bilingual, 16-field schema
$1.63
total spend (~$3.32 / video-hour)
3 tasks
washing Β· box assembly Β· fan assembly
β–Ά Open the sample01 studio
What makes it different

Four ideas behind the numbers

Route, don't prompt

Cheap 7B models have a hard perception ceiling that prompting can't lift (proven across 3 model families). Our M13 cascade routes the easy clips to free models and only buys frontier inference on the hard ones.

Cost is in the thinking, not the pixels

Measured: Gemini-3 video input is cheap; the spend is output + hidden reasoning tokens. Tuning that knob cut annotation cost 57% with zero quality loss.

Granularity is a constraint, not a budget

A segment-length cap + per-cycle rule fixes lazy under-segmentation of long clips far better β€” and cheaper β€” than paying for more model thinking.

Two-pass QA you can trust

A second independent model pass flags only the disagreements for human review, turning a black-box API call into an auditable, confidence-scored deliverable.