Google DeepMind's Gemini Robotics 1.5 makes robots think before they act, and puts its planning model in developers' hands

Google DeepMind on 25 September 2025 released two robot models that split the job of a general-purpose robot into a planner and a doer. Gemini Robotics 1.5 is a vision-language-action (VLA) model that turns camera images and instructions into motor commands and, according to the company, writes out its reasoning before it acts. Gemini Robotics-ER 1.5 is a vision-language model built for embodied reasoning that plans multi-step tasks, calls digital tools such as Google Search and hands steps to a VLA.

The access terms differ sharply. Gemini Robotics-ER 1.5 went live the same day as a preview in Google AI Studio and the Gemini API under the model identifier gemini-robotics-er-1.5-preview, the Google developers blog said. Gemini Robotics 1.5, the model that actually moves robot arms, is available only to select partners, DeepMind's announcement said. Neither model's weights were published.

The clearest number in the release sits in the accompanying technical report, a preprint that was not peer reviewed and was posted to arXiv on 2 October 2025 (arXiv:2510.03342). When Gemini Robotics-ER 1.5 orchestrated Gemini Robotics 1.5 on long-horizon tasks, total subtask failures were 22%, against 44.5% when Gemini 2.5 Flash acted as the orchestrator. Planning failures fell to 9% from 25.5%, action failures to 9% from 13% and success-detection failures to 4% from 6%, the report said.

The report also says the thinking VLA on its own reached a progress score of up to 44% on its hardest multi-step evaluations, while the full agent "frequently achieves scores near 80%". DeepMind said Gemini Robotics-ER 1.5 was evaluated on 15 academic embodied reasoning benchmarks, among them ERQA and Point-Bench, and claimed state-of-the-art results on aggregate. The report compares it with GPT-5, GPT-5-mini, Gemini 2.5 Pro and Gemini 2.5 Flash, but presents most per-benchmark results only in charts rather than in text, so MEKAWIRE does not reproduce individual scores here.

The second technical claim is motion transfer. DeepMind trained Gemini Robotics 1.5 on data from three quite different machines: the ALOHA 2 bimanual research platform, a bi-arm Franka setup and Apptronik's Apollo humanoid. The company says tasks learned on one of these bodies "also just work" on the others without specialising the model to each. The report contrasts this with its own Gemini Robotics On-Device models, which need a separate checkpoint for every embodiment.

The 1.5 generation is the third robotics release from DeepMind in 2025. The first Gemini Robotics and Gemini Robotics-ER models, built on Gemini 2.0, were announced on 12 March 2025 with Apptronik as hardware partner and Agile Robots, Agility Robotics, Boston Dynamics and Enchanted Tools as trusted testers. Gemini Robotics On-Device followed on 24 June 2025, offered to trusted testers through an SDK and adaptable to new tasks with as few as 50 to 100 demonstrations, according to DeepMind. Google also took part in Apptronik's $350 million Series A in February 2025, Apptronik said.

The competitive field has moved fast. Figure AI introduced Helix on 20 February 2025, pairing a 7-billion-parameter vision-language model running at 7 to 9 Hz with an 80-million-parameter control policy running at 200 Hz on embedded GPUs inside its robots. NVIDIA released Isaac GR00T N1 on 18 March 2025 as a downloadable 2-billion-parameter model under an NVIDIA licence that restricts commercial use. Physical Intelligence published pi0.5 on 22 April 2025 and released its base weights in its openpi repository in September 2025. Figure closed a $1 billion Series C at a $39 billion post-money valuation on 16 September 2025, nine days before DeepMind's release.

The idea behind the thinking VLA is simple to state. A conventional VLA maps pixels and a command straight to joint movements. Gemini Robotics 1.5 first produces a short internal monologue in language, breaking an instruction such as sorting laundry by colour into steps, and only then emits actions. DeepMind argues this makes behaviour easier to inspect and more robust to multi-step tasks. Gemini Robotics-ER 1.5 adds a tunable thinking budget, set through the thinking_config parameter, so a developer can trade latency for accuracy on harder spatial questions.

For robot makers, the split matters commercially. A company with its own controllers can call Gemini Robotics-ER 1.5 today as a cloud reasoning layer for pointing, object localisation, progress estimation and task planning, paying for API use rather than training a large model. The action model that would replace their own control stack remains behind a partner gate, so for now DeepMind sells the brain's planning cortex, not its motor cortex.

Safety remains an open question. DeepMind released an upgraded ASIMOV benchmark for semantic safety and says Gemini Robotics-ER 1.5 improves on its predecessor, but the report gives these results in charts and the tests are DeepMind's own. Cloud latency, pricing at scale and how motion transfer performs on bodies outside the three training platforms are not addressed by public data. The signals to watch are wider partner access to Gemini Robotics 1.5, a move of the ER model from preview to general availability, and third-party benchmark replications.