Gemini Robotics 2.5: 96% in Mercedes parts pilot, DeepMind says

Google DeepMind introduced Gemini Robotics 2.5 on November 10. The company reported a 96% success rate for Apptronik’s Apollo 2 humanoid in a parts-kitting proof of concept on Mercedes-Benz production lines.

The Hungary and Germany trial sorted metal parts from bulk bins into assembly-sequenced trays. DeepMind did not disclose the total attempt count or detailed scoring rules.

Robotics 2.5 translates camera observations and language instructions into movement. Its product page describes control ranging from dexterous hands and grippers to an entire humanoid body. It can adjust balance, step and bend to reach objects, extending manipulation beyond an arm operating from a fixed position. Google also describes behavior that adapts to unexpected situations and changing conditions.

The existing Gemini Robotics ER 2 handles higher-level reasoning. It interprets space, plans a sequence and hands execution to a lower-level action model. Google describes ER 2 as thinking about the next step while the robot moves, detecting completion and deciding when to retry. Planning and motor execution have distinct roles. ER 2 also supports robots coordinating a shared task according to their different capabilities.

Turning camera scenes into practice environments

DeepMind said Gemini 4 Argon acted as a simulation engine during development of 2.5. An October 13 developer article explains the underlying method: reconstruct objects and space from robot camera footage, then repeatedly adjust the simulated scene until its appearance and physical contacts match the recorded episode.

The developer workflow checks both visual alignment and the timing and magnitude of fingertip forces. Shape, mass and friction are among the adjustable properties. It also describes reviewing logs and camera footage after failures, diagnosing causes and checking fixes in a dry run before further physical action.

The public Asimov evaluation materials describe adversarial, multi-turn requests directed at robotics systems. The linked public evaluation card describes multi-turn tests of up to 10 steps, measuring attack success and excessive refusals of benign requests. Other tests cover payload and reach constraints, and uncertainty when gauges are blurred or obscured. The harness also describes estimating human proximity from images and choosing a stop or continue tool call. The documentation is readable publicly, while access to dataset files requires agreement to share Hugging Face account information.

These tests cover a different layer from physical safeguards. The linked October Gemini 4 safety report explicitly limits its robotics evaluations to high-level semantic safety reasoning and orchestration. The tests examine refusal of unsafe instructions, requests for clarification and stopping when a person is within a one-meter safety zone. Physical deployment also requires validation of functional safety, including certified hardware components, redundant mechanisms and real-time guarantees, the report says. It also reports that safety behavior was highly sensitive to the evaluation harness, including whether refusal was explicitly offered.

Apptronik separately describes hardware safety zones for Apollo 2. Its impact zone is designed to pause movement when an object enters the defined radius. A configurable perimeter governs behavior as people, equipment or obstacles approach. A chest display, lighting and speech communicate the robot’s state and task progress to people nearby. The hardware platform also offers modular bipedal or wheeled mobility configurations.

Robotics 2.5 is rolling out to research partners, and its model page directs prospective users to an early-access waitlist. The separate ER 2 page lists a public preview in Google AI Studio and the Gemini API. Those pages describe different access routes for the action model and the existing reasoning model.