Gemini Robotics ER 2: Release, Architecture, Benchmarks, Pricing, and Comparison Guide (2026)

Gemini Robotics ER 2 is Google DeepMind's newest embodied reasoning model, launched July 30, 2026, that acts as the "high-level brain" for robots — processing video, language, and sensor input to plan multi-step tasks, coordinate multiple robots, and hand off motor commands to a separate action model. It replaces Gemini Robotics-ER 1.6 as DeepMind's most capable reasoning layer for physical AI, and is publicly available today through the Gemini API, Google AI Studio, and in private preview on the Gemini Enterprise Agent Platform.

This guide covers exactly what changed, how it performs, what it costs, and how it stacks up against Gemini Robotics ER 1.6 and the rest of the Gemini Robotics model family.

This article covers:

1. What Gemini Robotics ER 2 is and its role as a high-level reasoning brain for robots

2. The difference between vision-language (VLM) and vision-language-action (VLA) models

3. The Gemini Robotics ecosystem: ER 2, Robotics 2, and On-Device 2

4. Key upgrades vs. ER 1.6 (continuous video, real-time streaming, multi-robot collaboration)

5. How ER 2 tracks task progress and detects failures using live video input

6. Architecture and orchestration: how ER 2 connects with tools, APIs, and robot control systems

7. Performance benchmarks, speed improvements, and safety enhancements

8. Pricing structure and cost comparison with ER 1.6

9. Real-world use cases, including multi-robot coordination and Boston Dynamics demos

10. How to get started, including SDK, API access, and developer workflow

What Is Gemini Robotics ER 2?

Gemini Robotics ER 2 is a vision-language model (VLM), not a vision-language-action (VLA) model — meaning it doesn't directly output motor commands. Instead, according to Google DeepMind's official launch post, it functions as an orchestrator: it observes a scene through continuous video, reasons about what needs to happen next, communicates with humans in natural language, and delegates physical execution to a lower-level VLA model or robotics API.

It's one of three models Google DeepMind released together on July 30, 2026:

• Gemini Robotics 2 — the vision-language-action (VLA) model that directly controls motors, is now capable of a whole-body humanoid control.

• Gemini Robotics ER 2 — the embodied reasoning model covered in this guide, is available now on Google AI Studio.

• Gemini Robotics On-Device 2 — an efficient VLA model built to run locally on robot hardware without an internet connection, is currently limited to early-access partners.

What's New in Gemini Robotics ER 2 vs. ER 1.6?

Google's developer blog post introducing the model, authored by DeepMind engineers Steven Hansen and Peng Xu, frames ER 2 as a "significant upgrade" over Gemini Robotics-ER 1.6 in three specific areas:

1. Continuous video understanding. ER 1.6 primarily reasoned over static snapshots. ER 2 watches continuous video feeds, allowing it to track its own task progress and catch failures — spills, slips, misalignments — mid-execution rather than only at the end.

2. Real-time streaming via the Live API. ER 2 integrates with the Gemini Live API through a bidirectional streaming endpoint built for latency-sensitive tasks, eliminating the "stop-and-think" pauses that limited earlier reasoning models.

3. Multi-robot collaboration. For the first time, different robot types can communicate through a shared semantic understanding and split a task based on each robot's physical strengths — something neither ER 1.6 nor the original Gemini Robotics-ER supported.

Gemini Robotics ER 2 Architecture

Architecturally, Gemini Robotics ER 2 keeps the same core design philosophy DeepMind has used since the original Gemini Robotics-ER: it's a Gemini-based multimodal model specialized for spatial reasoning, task planning, and success detection, sitting above the motor-control stack.

Developers integrate it by declaring low-level control interfaces, such as VLA models or navigation APIs, as callable tools, then streaming multimodal video, audio, or text directly into the model. ER 2 can also natively call external tools, including Google Search, during its reasoning process. Google has published working examples of this setup, including a Boston Dynamics Spot integration where ER 2 orchestrates Spot's navigation and manipulator APIs to autonomously fetch an object on a spoken command.

Gemini Robotics ER 2 Benchmarks

DeepMind published specific, verifiable performance numbers alongside the launch — a level of transparency worth noting, since it lets developers judge the model on real evaluation data rather than marketing language alone.

Beyond these headline numbers, DeepMind also reports gains on general spatial reasoning: success/failure detection now runs on raw video instead of static frames, and instrument reading has been extended beyond simple dials to digital displays, linear scales, rulers, and liquid thermometers, tested across ten instrument types. Full methodology is available in DeepMind's Gemini Robotics 2 Safety Technical Report.

Gemini Robotics ER 2 Price and Gemini Robotics ER 1.6 Price

Unlike some enterprise robotics platforms that require a sales conversation, Gemini Robotics ER 2 is priced transparently on Google's official Gemini API pricing page — the same page that lists every other Gemini model.

A few practical notes on pricing:

• Both ER 2 and ER 1.6 offer a free tier through Google AI Studio for development and testing, with the usual caveat that free-tier content may be used to improve Google's products.

• Batch API processing cuts the cost roughly in half for both models — useful for offline evaluation, training-data labeling, or non-real-time robotics research.

• ER 2 costs exactly double ER 1.6 on both input and output — a straightforward tradeoff to weigh against ER 2's video-understanding and multi-robot capabilities, which ER 1.6 simply doesn't support.

• Pricing for the full VLA models — Gemini Robotics 2 and Gemini Robotics On-Device 2 — has not been published, since both remain limited to early-access partners rather than open API access.

Gemini Robotics ER 2 vs. Gemini Robotics ER 1.6: Full Comparison

Bottom line: if your application needs live video monitoring, multi-robot coordination, or real-time streaming responsiveness, ER 2 is the clear upgrade. If you're running simpler, single-robot, non-time-critical reasoning tasks, ER 1.6 remains available at half the token cost.

Gemini Robotics On-Device: The Offline Counterpart

Gemini Robotics On-Device 2 is the sibling model built specifically to run locally on robot hardware, without relying on cloud connectivity that is critical for factories, warehouses, or field robots operating with limited or no internet access. According to DeepMind, it can adapt to entirely new dual-arm robot embodiments using fewer than 200 training examples collected over just a few hours, inheriting the "motion transfer" technique first introduced in Gemini Robotics 1.5. Unlike Gemini Robotics ER 2, the on-device model is currently restricted to early-access partners rather than open API access.

Gemini Robotics SDK: Building on Your Own Hardware

The Gemini Robotics SDK is Google's developer toolkit for testing and adapting Gemini Robotics models to specific robots and tasks. Originally released alongside the first Gemini Robotics On-Device model, the SDK supports evaluation in Google's MuJoCo physics simulator and allows developers to fine-tune model behaviour for new robot embodiments using as few as 50–100 demonstrations.

For Gemini Robotics ER 2 specifically, Google has published ready-to-use Jupyter notebook examples on GitHub, including the full Boston Dynamics Spot demo, making it one of the more accessible entry points into physical AI development for teams that already work with the Gemini API.

Gemini Robotics Paper: The Research Behind the Models

For readers who want the underlying research rather than the product announcement, DeepMind has published a peer-reviewable technical report for each major generation:

• "Gemini Robotics: Bringing AI into the Physical World" (arXiv:2503.20020, March 2025) — the foundational paper introducing the original Gemini Robotics and Gemini Robotics-ER models, built on Gemini 2.0.

• "Gemini Robotics 1.5: Pushing the Frontier of Generalist Robots" (arXiv:2510.03342) — introduces the Motion Transfer mechanism and "thinking before acting" reasoning that ER 2 builds on.

• Gemini Robotics 2 Safety Technical Report — the safety-specific evaluation accompanying the July 2026 release.

A dedicated technical report specifically for the July 2026 Gemini Robotics 2 / ER 2 generation had not been published as a standalone arXiv paper at the time of writing; the safety report and official blog posts are currently the primary documented sources.

Real-World Example: Boston Dynamics' Spot Fetches a Snack

To demonstrate ER 2's tool-orchestration capabilities in practice, DeepMind partnered with Boston Dynamics to show Gemini Robotics ER 2 commanding Spot's navigation and manipulator APIs entirely through natural language.

Separately, Apptronik's Apollo 2 humanoid, running the companion Gemini Robotics 2 VLA model, demonstrated whole-body coordination — walking to a table, picking up a watering can, and placing it precisely on a shelf, all orchestrated by ER 2's high-level planning.

How to Get Started with Gemini Robotics ER 2?

1. Get a free API key through Google AI Studio.

2. Review the Gemini API robotics documentation for spatial reasoning, agentic vision, and task orchestration guides.

3. Clone the official robotics-samples GitHub repository for working notebook examples.

4. For enterprise deployments, request access to the Gemini Enterprise Agent Platform private preview.

FAQs

What is Gemini Robotics ER 2?

It's Google DeepMind's most capable embodied reasoning model, launched July 30, 2026, that serves as a robot's high-level decision-making layer — interpreting video and language, planning multi-step tasks, and coordinating with other robots, while handing off actual movement to a separate action model.

What is the Gemini Robotics ER 2 price?

Through the official Gemini API, standard pricing is $2.00 per million input tokens and $10.00 per million output tokens, with a free tier available through Google AI Studio and roughly 50% lower rates on the Batch API.

What is the Gemini Robotics ER 1.6 price?

Gemini Robotics-ER 1.6 is priced at $1.00 per million input tokens (text/image/video), $2.00 for audio input, and $5.00 per million output tokens — half the cost of ER 2.

How is Gemini Robotics ER 2 different from Gemini Robotics ER 1.6?

ER 2 adds continuous video understanding for real-time progress tracking, bidirectional low-latency streaming via the Gemini Live API, and multi-robot collaboration — none of which ER 1.6 supports. It also scores notably higher on safety and spatial-reasoning benchmarks.

What is Gemini Robotics On-Device?

It's a separate, efficient vision-language-action model designed to run locally on robot hardware without an internet connection, currently available only to early-access partners rather than the general public.

What is the Gemini Robotics SDK?

A developer toolkit that lets teams evaluate and fine-tune Gemini Robotics models on their own tasks and robot embodiments, using Google's MuJoCo simulator and, for ER 2 specifically, published GitHub notebook examples.