Gemini Robotics 2.5 accelerates real-world general dexterity
Google DeepMind announced Gemini Robotics 2.5, described as its most capable vision-language-action model to date, with the release credited to the Gemini Robotics team led in the announcement by Carolina Parada. The company says the model is now rolling out to its robotics partners. The pitch is breadth and speed of learning: Gemini Robotics 2.5 is meant to let robots pick up a wider range of skills faster, improving dexterity, reliability and safety enough for unstructured real-world settings rather than controlled lab environments.
A vision-language-action model is the layer that connects perception and reasoning to motor commands, so a single model can look at a scene, interpret a spoken or written task, and drive arms and hands. DeepMind says Gemini Robotics 2.5, like its predecessors, works natively with multiple robot types. The two named platforms are Apptronik's Apollo 2 humanoid, fitted with a five-finger hand the announcement calls SharpaWave with 22 degrees of freedom, and the Franka Duo platform, which uses two-fingered parallel grippers.
The evaluation covers more than 200 distinct tasks, a set DeepMind says could be scaled further. Examples cited include serving food with tongs, tying a bow on a gift box, and pipetting a liquid, chosen to combine tool use, manipulation of deformable objects and work with delicate items. Against previous baselines, the company reports 30% to 100% higher success rates depending on task type, and completion 20% to 30% faster. Those figures are aggregate ranges rather than per-task numbers, so a developer reading them should note the lower bound matters most: the weakest task categories still gain substantially, but the spread suggests performance remains uneven across skills.
The most concrete real-world result comes from a proof of concept with Apptronik and Mercedes-Benz on active automotive production lines in Kecskemét, Hungary, and Sindelfingen, Germany. Apollo 2 robots performed automotive kitting: sorting individual parts from bulk bins into organized trays in the exact order each vehicle on the line requires. DeepMind says the system achieved a 96% success rate autonomously. The architecture splits the job in two. An embodied reasoning model, Gemini Robotics ER, tracks the workflow, while the Gemini Robotics 2.5 actions model controls the hands and arms, reaching into bins, picking up metal brackets and placing them in sequencing containers while correcting for gaps.
Apptronik co-founder and CEO Jeff Cardenas said in the announcement that Apollo was built for useful work alongside people in real industrial settings, and that pairing Apollo 2 with Gemini Robotics produced dramatic improvement within a few weeks of post-training, carrying over to Mercedes-Benz production with very little drop-off. This suggests the practical story is fine-tuning efficiency as much as raw model capability: a short post-training cycle on top of a general model, transferring into a live factory. A likely trade-off is that a 96% success rate on a repetitive kitting task is impressive but still short of the near-perfect reliability factories typically demand, which is why the framing is a pilot and a proof of concept rather than a deployment claim.
Safety gets two distinct treatments. For physical safety, Gemini Robotics 2.5 uses inference-time reasoning guidance to steer around hazards in real time, which DeepMind likens to safety classifiers layered around Gemini's responses. The company reports this lifted safe task completion by 33% around obstacles and 43% around nearby people. For agentic safety, DeepMind says Gemini 4 Argon currently leads its agentic robotics evaluations on unsafe task refusal, resolving uncertainty and working near humans, and it is releasing ASIMOV-Jailbreak, an open-source benchmark with hundreds of prompts disguising unsafe requests across dozens of scenes. Argon is reported at 99.5% jailbreak resistance versus 82% for the previous baseline, Gemini Robotics ER 2. The benchmark's release is arguably as useful to the field as the score itself, since a shared adversarial test lets outside teams compare safety claims.
Gemini 4 Argon also underpins the development story. DeepMind says researchers need large volumes of physical data and efficient algorithms, and that Argon acted as a simulation engine to ease both bottlenecks. From a real-world camera scene it can generate thousands of varied simulated practice runs and build a realistic digital twin; combined with real demonstrations, this helps the robot generalize to new scenes and objects. Argon can also help hypothesize why failures occurred from video footage and test candidate fixes in simulation and on real robots. In practice, simulation-to-real transfer of this kind is a well-known way to cut the cost of collecting physical demonstrations, though the gap between simulated and real conditions remains the usual place these pipelines break down.
There is no publicly stated pricing or general availability beyond the partner rollout, and the reported figures are DeepMind's own evaluations rather than independent results.
Why it matters: Teams building on humanoids and industrial arms now have a VLA model pitched at multi-robot compatibility, faster task learning and measurable safety steering, with a named production-line pilot behind it. If the reported gains hold outside DeepMind's evaluations, the practical shift is shorter post-training cycles before a robot is useful on a factory floor, plus a new open benchmark for jailbreak resistance that safety reviewers can test against. Inference: expect near-term pressure on robotics teams to report success rates and safety metrics in comparable terms, and to weigh simulation-engine tooling as part of the training pipeline rather than an afterthought.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Our most capable vision-language-action model enhances robot dexterity, reliability, and safety