Physical Tool‑Use Capabilities of Gemini 3.7 Flash

August 17, 2026

Gemini 3.6 Flash and Gemini 3.7 Flash were each given control of YAM arms to complete the following prompt: “Lift the towels with both your arms high enough to then drop it to cover the block.”

Gemini 3.7 Flash scores 92% on this task, up from 32% for Gemini 3.6 Flash, released just a month earlier.

Physical Tool‑Use — Covering a Block with a Towel

2026-08-14T16:08:39.394579 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/ 0 10 picks towel 20 both arms 40 lifts up 60 partial cover 80 full cover 100 judges correctly Gemini 3.6 Flash Gemini 3.7 Flash 32 92 hit 100-call budget robocurve.org

Mean rubric score across five runs per model (max 100).

Milestones

Each run is graded by hand on a cumulative milestone ladder.

picks towel

1. picks towel 10 pts

uses both arms

2. uses both arms 10 pts

lifts towel up

3. lifts towel up 20 pts

partial / full cover

4. partial / full cover 20 pts / 40 pts

A sixth milestone, judges correctly (20 pts), credits a final done / give‑up verdict that matches what actually happened on the table.

Per-Turn Latency

Lower is better (faster inference).

2026-08-15T12:15:50.911471 image/svg+xml Matplotlib v3.11.1, https://matplotlib.org/ 0 5 10 15 20 25 Gemini 3.7 Flash 9.0 s Gemini 3.6 Flash 17.5 s seconds per turn robocurve.org

Mean inference latency per turn. Robot motion excluded.

All Runs

ModelScoreTime (min)TranscriptRerun (.rrd)
Gemini 3.7 Flash1002.3viewdownload
Gemini 3.7 Flash1002.8viewdownload
Gemini 3.7 Flash1002.8viewdownload
Gemini 3.7 Flash1003.3viewdownload
Gemini 3.7 Flash604.1viewdownload
Gemini 3.6 Flash5018.2viewdownload
Gemini 3.6 Flash4032.0viewdownload
Gemini 3.6 Flash4034.6viewdownload
Gemini 3.6 Flash3022.6viewdownload
Gemini 3.6 Flash037.2viewdownload

Technical Specifications

ArmsI2RT YAM bimanual arms (two 6‑DoF arms with parallel grippers)
Camerasfixed overhead Intel RealSense D435 + one Intel RealSense D405 on each wrist, 224×224 RGB
ModelsGemini 3.6 Flash and Gemini 3.7 Flash on the Interactions API, effort high, 5 runs each
Instruction“Lift the towels with both your arms high enough to then drop it to cover the block.”
Max turns100 LLM calls per run
Actionsone move per LLM call (action horizon 1); arm speed capped at 10% for safety
Harnessinspect‑robots 0.53.1; collision guardrail off, joint‑effort reporting on, rig facts / formulas / advice documents in the system prompt; full wire capture
Scoringmanual milestone grading from the run videos