Gemini Robotics-ER 1.6
Google DeepMind, US · April 2026
- Parameters
- not published
- No parameter count published.
- GPU memory
- not published
- weights at bf16, computed
- Inference latency
- not published
- per action step
- Weights
- closed
- closed, available through the Gemini API
What it is
Gemini Robotics-ER 1.6 is an incremental embodied reasoning release from April 2026 whose distinguishing feature is reading industrial instruments, where it reaches 86 percent accuracy against 23 percent for ER 1.5 and 67 percent for Gemini 3.0 Flash, rising to 93 percent with agentic vision. It also improves pointing, counting, multi-view success detection and safety compliance. It is the first ER model with a named production deployment, running on Boston Dynamics Spot for facility inspection. The caveats are that the headline benchmark is internal to Google and that the model is already slated for deprecation at the end of August 2026 in favor of ER 2.
Architecture
- Backbone
- Vision-language model for embodied reasoning. The release compares it against Gemini 3.0 Flash as the general-purpose baseline.
- Action head
- None. It outputs pointing, localization, counting, success detection, instrument readings and plans in text.
- Parameters
- No parameter count published.
- Pretraining data
- Not published.
- Embodiments
- Boston Dynamics Spot (deployed for facility inspection and instrument monitoring)
What hardware it needs
This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.
The authors publish no memory or latency figure for this model. Everything above is computed from the parameter count.
Reported results
Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.
Simulation
No results in this category are published for this model.
Real world
Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.
- Instrument reading (Google DeepMind internal) industrial gauges and sight glasses with agentic vision93%accuracysource
- Instrument reading (Google DeepMind internal) industrial gauges and sight glasses86%accuracyGemini Robotics-ER 1.5 scores 23 percent and Gemini 3.0 Flash 67 percent on the same task. This is a first-party internal benchmark, not an academic suite.source
Fine tuned tasks
No results in this category are published for this model.
On real hardware
Deployed inside Boston Dynamics Spot for facility inspection and instrument monitoring, which is the first Gemini Robotics-ER model with a named production deployment. Its distinguishing capability is reading industrial instruments such as gauges and sight glasses, at 86 percent accuracy or 93 percent with agentic vision.
Fine tuning it yourself
No fine-tuning. The suggested improvement path for a specialized application is to submit 10 to 50 labeled images that show the failure modes you care about.
Where it helps, where it does not
Strengths
- Instrument reading jumps from 23 percent (ER 1.5) to 86 percent, or 93 percent with agentic vision
- Named real deployment on Boston Dynamics Spot for facility inspection
- Improved multi-view understanding for coordinating several camera streams
- Better safety compliance on adversarial tasks, reported as plus 6 points on text and plus 10 on video for injury risk perception versus Gemini 3.0 Flash
Limits
- The headline instrument-reading number is a first-party internal benchmark with no external replication
- Closed weights, API only
- Reasoning model only, it cannot produce motor commands
- Already scheduled for deprecation at the end of August 2026, roughly four months after release, which makes it a poor foundation to build on
- No parameter count, latency or pricing published in the announcement
Sources
Everything on this page was taken from these documents. Where they disagree with what you read here, they win.
- https://deepmind.google/blog/gemini-robotics-er-1-6/
- https://ai.google.dev/gemini-api/docs/robotics-overview
- https://blog.google/innovation-and-ai/models-and-research/google-deepmind/gemini-robotics-er-1-6/
Entry last checked 2026-08-11.
Try a policy on a real arm
A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.