What Gemini Robotics-ER 1.6 Is
Gemini Robotics-ER 1.6 is Google DeepMind's upgraded "reasoning-first model" for robots. Released April 14, 2026, it specializes in embodied reasoning—reasoning that assumes a body in the world: cameras, grippers, joints, gravity, friction, occlusions.
"ER" = Embodied Reasoning. The model doesn't just see images. It understands spatial constraints, physical limits, and what happens when objects interact.
Availability
- Gemini API (developer access)
- Google AI Studio (https://aistudio.google.com/prompts/new_chat?model=gemini-robotics-er-1-6-preview)
The Benchmark That Matters
Instrument reading accuracy: 23% → 93%
The previous model (ER 1.5) could barely read analog gauges. ER 1.6 nails it. That's a 4x improvement.
| Model | Instrument Reading Success |
|---|---|
| Gemini Robotics-ER 1.5 | 23% |
| Gemini 3.0 Flash | 67% |
| Gemini Robotics-ER 1.6 | 86% |
| ER 1.6 + Agentic Vision | 93% |
This isn't a marginal improvement. It's a capability unlock. Industrial robots can now patrol facilities and read pressure gauges, sight glasses, thermometers—things most computer vision systems still fumble.
How It Actually Works
Agentic Vision Approach:
- Zoom into gauge images to resolve fine detail
- Point to identify dial positions
- Execute code—synthesizes Python scripts, runs CV/math tasks
- Apply world knowledge to derive final reading
The model treats vision as stepwise reasoning. It doesn't just "recognize" a gauge. It zooms, points, calculates proportions, and cross-references with what it knows about how gauges work.
Other Improvements
| Capability | ER 1.5 | ER 1.6 |
|---|---|---|
| Object counting (overlapping) | Hallucinations | Correct |
| Success detection | Single view | Multi-view synthesis |
| Pointing to absent objects | Often hallucinated | Correctly declines |
| Safety compliance | Baseline | Superior on adversarial tasks |
Real-World Deployment: Boston Dynamics Spot
Boston Dynamics integrated ER 1.6 into Orbit AIVI-Learning. Spot robots can now:
- Patrol industrial facilities autonomously
- Read data from instruments during inspection rounds
- Detect puddles, count pallets, verify lever positions
Marco da Silva (VP/GM Spot, Boston Dynamics):
"Capabilities like instrument reading and more reliable task reasoning will enable Spot to see, understand, and react to real-world challenges completely autonomously."
The Competitor Landscape
| Competitor | Focus | Key Differentiator |
|---|---|---|
| NVIDIA Isaac | Full robotics stack | Isaac Lab simulation + Jetson Thor edge deployment |
| Tesla Optimus | Physical humanoid | Tesla Autopilot/Dojo training |
| Figure 02 | Deployed humanoid | OpenAI partnership, BMW facility deployment |
| Apptronik Apollo | Humanoid platform | Gemini Robotics integration partner |
ER 1.6 is a reasoning layer. NVIDIA Isaac is a platform stack. Different bets on where value accrues in robotics.
The Reality Check
From Hacker News (216 points, 82 comments):
Top concern: Latency.
"The gauge-reading example is great, but having the system synthesize Python scripts, run CV tasks, come back with the answer is currently quite slow."
ER 1.6 is a high-level orchestrator. Real-time closed-loop motor control still requires careful architecture decisions. It's not designed for safety-critical tasks where milliseconds matter.
Known Limitations
- Latency: Inference too slow for real-time control loops
- Edge deployment: No announced on-prem/edge option
- Robustness: Benchmark conditions vs dusty/poor lighting environments
- Power: Cloud-based inference requires connectivity
What This Means
Industrial inspection robots just got smarter. Not "general purpose humanoid" smarter—but specifically better at reading analog instruments, counting overlapping objects, and knowing when to say "I can't see that."
That's the quiet breakthrough. Most robot vision systems hallucinate when asked to point to things they can't actually see. ER 1.6 learned to decline. That's not a benchmark number. That's a behavior that prevents failures.
Sources
- DeepMind Official Blog: https://deepmind.google/blog/gemini-robotics-er-1-6/
- Boston Dynamics Integration: https://www.therobotreport.com/boston-dynamics-and-google-deepmind-are-using-gemini-to-make-spot-smarter/
- IEEE Spectrum Coverage: https://spectrum.ieee.org/boston-dynamics-spot-google-deepmind
- Hacker News Discussion: https://news.ycombinator.com/item?id=47779094