Gemini Robotics-ER 1.6: 身体化推論を強化し実世界ロボット制御を加速 Gemini Robotics-ER 1.6: Powering real-world robotics tasks through enhanced embodied reasoning

Google DeepMind has released Gemini Robotics-ER 1.6, the latest version of its embodied reasoning model designed to bring Gemini's multimodal capabilities to physical robots. The update, accessible to developers through the Gemini API, focuses on improving spatial understanding and multi-step task planning for real-world robotic deployments.

The ER (Embodied Reasoning) line extends Gemini's general-purpose vision and language abilities into the robotics domain. Given camera input, the model can localize objects, infer affordances such as where and how an item can be grasped or manipulated, and generate sequences of actions to accomplish a high-level goal. According to Google DeepMind, version 1.6 delivers measurable gains on spatial-reasoning benchmarks, handles longer-horizon tasks more reliably through better decomposition, and exhibits stronger judgment on safety-relevant decisions compared with its predecessor.

In typical robot control stacks built around these models, the ER component effectively functions as the high-level "brain," interpreting instructions, perceiving the scene, and planning steps, while a separate vision-language-action (VLA) policy or low-level controller translates those plans into motor commands. Gemini Robotics-ER 1.6 appears to be positioned squarely in that upper layer, intended to be paired with action models or conventional controllers tuned to specific hardware. This separation of concerns has become a common pattern as developers seek to combine the broad world knowledge of large multimodal models with the real-time precision required for physical actuation.

The release fits into a broader push by Google DeepMind across the robotics stack this year. The company has previously introduced a cloud-based Gemini Robotics model, an on-device variant aimed at lower-latency or offline use cases, and earlier iterations of the reasoning-focused Gemini Robotics-ER family. Taken together, the lineup signals an intent to offer a tiered set of foundation models that span from heavy-duty server inference to constrained embedded deployments.

The competitive landscape around general-purpose robot foundation models has intensified sharply since 2024. Figure AI has been promoting its Helix system, Physical Intelligence has iterated on its pi-series policies, NVIDIA continues to expand the GR00T humanoid platform, and OpenAI has reportedly re-entered robotics after winding down its earlier effort. Several of these projects converge on a similar architectural bet: a large multimodal or VLA model serving as the cognitive core, coupled with learned or classical low-level control. The race to bring such foundation models into physical environments has become one of the more visible frontiers in applied AI.

Research context for ER-style models draws on publicly available datasets and benchmarks such as Open X-Embodiment, which aggregates demonstrations across many robot embodiments, as well as open VLM efforts that have explored how language-and-vision pretraining transfers to embodied tasks. By exposing Gemini Robotics-ER 1.6 through an API, Google DeepMind effectively lowers the barrier for startups and academic groups to plug the model into their own hardware platforms and evaluate it on bespoke tasks, without needing to train a comparable backbone from scratch.

Practical deployment, however, is likely to remain non-trivial. Latency is a recurring concern when a cloud-hosted reasoning model sits in the perception-to-action loop, and safety guarantees are difficult to obtain from probabilistic planners operating in unstructured environments. Generalization across unseen objects, layouts, and embodiments also continues to be an open research problem; benchmark gains do not always translate cleanly to robust behavior on a physical robot. Google DeepMind highlights improvements in these areas, but independent evaluation by external developers will be important to gauge how far the new model moves the needle.

For now, Gemini Robotics-ER 1.6 represents an incremental but notable step in the company's robotics roadmap. Whether it becomes a default high-level reasoning layer for third-party robot builders may depend on how it compares in practice with rival foundation models, on pricing and rate limits via the API, and on how easily it integrates with the diverse VLA policies and control stacks already in use across the industry.