Google DeepMind unveiled Gemini Robotics 2 on July 30, 2026, with three sub-models, but only gemini-robotics-er-2" class="entity-chip">Gemini Robotics ER 2 is publicly available. The embodied reasoning VLM hits nearly 60 percent video frame completeness, up from the 1.6 release.
Key facts
- Gemini Robotics 2.0: three sub-models, only ER 2 public
- ER 2 achieves ~60% video frame completeness accuracy
- ER 2 processes live video from robot cameras
- Released via Gemini Live API, July 30, 2026
- Dexterity and safety models not yet released
Google DeepMind's Gemini Robotics 2.0, announced July 30, 2026, brings a trio of sub-models aimed at generalist robots — machines that can handle any task a human could, what DeepMind scientists call "physical AGI." According to Ars Technica, the release includes an upgraded "embodied reasoning" model, a dexterity model, and a safety model. Only the first, Gemini Robotics ER 2, is available to developers now via the Gemini Live API.
The headline capability is live video processing. ER 2, a vision language model (VLM), can now track progress from the robot's cameras in real time, allowing the system to adjust as it moves from step to step. Google claims ER 2 classifies video frame completeness with almost 60 percent accuracy — far from perfect, but better than the 1.6 release or competing models' visual understanding. The company did not disclose benchmark details or training costs.
What's Actually New
The dexterity and safety models remain unreleased, with Google staying silent on timelines. That's a notable gap: the press release touts "improved dexterity" and "safety," yet developers can only test the reasoning layer. The two withheld models are where the claimed leaps in humanoid hand control and safe operation would land. Google's decision to gate them suggests either they're not production-ready or the company is holding back its best robotics assets.
Why This Matters
This release lands as Google's AI capex hits record levels — the company posted its first negative free cash flow since 2004 in late July, per our reporting. Robotics is a logical outlet for that spend, but the 60 percent video accuracy figure is a sobering counterpoint. It's a reminder that embodied AI remains in early innings; the gap between demo videos of backflipping robots and reliable generalist operation is still wide. For developers, ER 2 via the Gemini Live API is a concrete tool to test, but the real prize — the dexterity and safety models — stays locked behind Google's roadmap.
What to watch
Watch for Google's release timeline for the dexterity and safety sub-models. If they arrive within two quarters, expect a push into humanoid robot control. Also track developer adoption of ER 2 via the Gemini Live API and any benchmark comparisons against competitors like AgiBot's WITA-Omni, which scored 85.21 on DailyOmni.
Source: arstechnica.com
