For an inspection-tray prototype, “Gemini Robotics or Cosmos?” is too broad a buying question. Your code needs a particular output: a text plan, generated development data, or numerical actions in the robot’s format.
Select that interface first, then confirm access and the machine that will run it. Both families cover several roles; the model name alone does not specify the integration.
What will your application consume?
| Exact offering | Interface and access |
|---|---|
| Gemini Robotics 2 | Vision/language → motor control for supported robots. Early-access partners; confirm embodiment and deployment terms. |
| Gemini Robotics ER 2 | Text/image/video/audio → text. Hosted AI Studio/Gemini API; Enterprise Agent Platform is a private preview. |
| Gemini Robotics On-Device 2 | Text/images/numerical robot state → numerical actions. Trusted testers; confirm local inference requirements. |
| Cosmos 3 Reasoner | Text/vision → text. Choose a supported serving integration and own its runtime. |
| Cosmos 3 Generator | Supported text/vision/sound/action inputs → vision, sound or actions. Select the model/integration for that output; Policy-DROID needs a compatible embodiment. |
Sources: Robotics 2 release , ER 2 card , On-Device 2 card , Cosmos 3 reference .
For numerical actions, verify dimensions, units and coordinate frames against the controller. A correctly formatted JSON array can still describe the wrong robot.
Check the application limits
The Cosmos3-Edge card describes physics limitations; ER 2’s terms exclude safety-critical applications. Review those constraints for the proposed application before treating either output as usable.
Budget the runtime you are choosing
ER 2’s API route avoids supplying model-serving hardware. On-Device 2’s public card does not give a general inference bill of materials; its TPU discussion concerns training. Its evaluated scope also does not establish whole-body mobile control.
Cosmos hardware recommendations vary: Super 64B includes H200/B200/GB200; Nano 16B, RTX PRO 6000/H100/B200; Edge 4B, Jetson AGX Orin/Thor/RTX PRO 6000. Check the model matrix for the selected variant.
Execution rate is not inference speed. NVIDIA reports about 1.53 seconds to generate 32 Edge actions played at 15 Hz: roughly 2.13 seconds of motion. That is not 15 model evaluations per second. Measure observation age and replanning delay in your controller. Deployment example .
Complete one component contract : exact version, access, input/output, hardware, action consumer and update owner. Leave unavailable access or unconfirmed controller support visible in the estimate before committing to a manipulation prototype.
Output documentation rechecked September 7, 2026. No integration or performance comparison was executed.
Sources and service screenshots (4)
Public reference pages captured September 6, 2026.
