Google DeepMind unveils Gemini Robotics 2.5, citing 96% success in Mercedes parts sorting
On November 10, 2026, DeepMind introduced Gemini Robotics 2.5 and described it as its most capable vision-language-action model so far. The model turns camera views and language instructions into arm and hand movements. DeepMind says it was tested on more than 200 tasks, including serving food with tongs and pipetting liquids, with success rates 30 to 100 percent above earlier baselines depending on task type. The same evaluations finished 20 to 30 percent faster, according to The Decoder's account of the announcement.
The system splits the work between two models. Gemini Robotics ER tracks the workflow and selects actions over longer horizons, while Gemini Robotics 2.5 controls the arms and hands. In a proof of concept with Apptronik, Apollo 2 humanoids worked on active Mercedes-Benz lines in Kecskemét, Hungary, and Sindelfingen, Germany, reaching into supply bins and sorting metal brackets into sequencing containers. DeepMind reports 96 percent success for that kitting work. Gemini 4 Argon, a separate Google model, served as the simulation engine during development.
DeepMind says the model uses inference-time reasoning guidance to steer around physical hazards in real time. The company also published ASIMOV-Jailbreak, an open evaluation built from adversarial prompts meant to push robots toward dangerous behavior. Rollout is limited to robotics partners, which DeepMind names as Agile, Apptronik and Mercedes-Benz. The release follows the Gemini Robotics 2 family that DeepMind introduced on July 30, 2026.
Why it matters
For robotics founders, the Mercedes result is a published reference point, but it needs qualifiers. DeepMind reports 96 percent success and has not said how many picks that figure covers. A DeepMind robotics leader reportedly acknowledged in an interview that moving from 96 percent to the 99.9 percent reliability common in industrial settings will be a steep climb. Teams selling into factories should expect buyers to ask for failure rates over time, not pilot scores.
For investors, the constraint is access. Rollout is limited to partners, and no public API or pricing has been disclosed, so third-party revenue depends on partner programs for now. Google also pairs a planning model, an action model, a simulator and a safety benchmark, which raises the bar for startups that own only one layer. Diligence should ask which layer a company controls and whether its safety claims can be checked against shared tests such as ASIMOV-Jailbreak.
Key points
- 1Gemini Robotics 2.5 was introduced on November 10, 2026 and is rolling out only to DeepMind's robotics partners.
- 2Success rates were 30 to 100 percent higher than earlier baselines across more than 200 evaluated tasks, depending on task type.
- 3Tasks finished 20 to 30 percent faster than before on the same evaluations, according to DeepMind's reported results.
- 4The Mercedes-Benz proof of concept reported 96 percent success on automotive kitting using Apptronik's Apollo 2 humanoid.
