Separate commercial, limited, prototype and development stages.
Robot product dossier · 12
Gemini Robotics 1.5 and ER 1.6
Google DeepMind · Developer access and selected partners · VLA and embodied reasoning models
Research / demonstrationResearch and developer-access models; access, hardware support and production readiness varyLast reviewed: 11 August 2026
Decision questionHow can a robotics team test whether a foundation model transfers beyond its published demonstrations?
12
FUURAA thesis
Gemini Robotics 1.5 and ER 1.6 are best treated as model-layer capabilities, not complete robot products. Their value appears only through hardware, sensing, control, safety and evaluation supplied by a wider system.
Reading method
Do not ask only what the robot can do; ask how the whole system works, fails and recovers.
Inventory software, payloads, facilities and people beyond the robot.
Move from demos to repeated tests, field data and auditable records.
State what first-party reporting still cannot prove.
System and workflow map
Four interdependent layers turn a product name into real operating conditions.
Each layer explains its role and what procurement, research or deployment teams should watch.
Multimodal reasoning
Models interpret language, visual context and task goals.
Embodied action
Model outputs must be grounded into actions supported by a robot.
Hardware adaptation
Transfer across embodiments requires calibration, data and interface work.
Safety wrapper
Independent constraints and monitors are needed around probabilistic outputs.
Decision checklist
Five checks turn product interest into verifiable procurement, research or pilot questions.
Define the task, interfaces, safety and operations before requesting evidence that supports a decision.
- 01
Task fit
Choose a task suite with seen, shifted and genuinely novel variants.
- 02
Integration load
Map cameras, robot APIs, control rate, compute, network and data governance.
- 03
Safety case
Keep safety-critical limits outside unconstrained model generation.
- 04
Operating model
Plan prompt, data, model-version, monitoring and fallback ownership.
- 05
Evidence to request
Request success distributions, latency, interventions, unsafe proposals and cross-robot results.
Scope boundaries
Product availability, a video demonstration or a commercial agreement each supports only bounded conclusions.
- 01
A model release or API does not constitute a deployable robot system.
- 02
Demonstrated transfer does not imply reliable performance on arbitrary hardware or tasks.
- 03
Access terms, supported embodiments and model behaviour may change; verify current documentation.
Sources and evidence status
Primary product material and independent evaluation context are kept distinct.
Dates, source identity, supported conclusions and reading boundaries remain reviewable.
01
Google DeepMind · 2026-04-14
↗02Gemini Robotics 1.5 and ER 1.6 — official product or programme source
Models for physical reasoning, planning, tool use and translating visual information and instructions into robot actions.
Google DeepMind · Accessed 2026-08-11
↗Open X-Embodiment: robotic learning datasets and RT-X models
Supporting material used to calibrate evaluation method and scope boundaries.