DeepMind details Gemini Robotics 2 powering Apptronik's Apollo 2 humanoid

Gemini Robotics 2 is DeepMind's push to bring frontier reasoning into physical embodiment, and the Apptronik Apollo 2 demos are the concrete proof point. In DeepMind's examples, the humanoid handles full-body autonomous functions with advanced real-time reasoning — parsing a natural-language instruction like 'Put the watering can into the green bin in the bottom shelf,' then walking to a table, locating the object and picking it up. The framing from DeepMind's VP is deliberately grand: rewriting the rules of robotics by grounding a general-purpose model in real-world manipulation.

Technically, the interesting shift is coupling a large multimodal reasoning model with low-level motor control tightly enough to execute multi-step physical tasks from open-ended commands — the long-standing hard problem in robotics being the gap between high-level intent and reliable actuation. Announced originally July 30, this week's blog deepened the technical narrative and the Apollo 2 partnership.

The competitive backdrop is a broadening 'physical AI' race that also showed up this week in AWS's NVIDIA Cosmos 3 model factory on SageMaker HyperPod and NVIDIA's edge-reasoning work. The skeptical read on humanoid demos is well-earned: curated single-task clips rarely reflect reliability across cluttered, unstructured environments, and 'walks to the table and picks up the object' is exactly the kind of demo that looks solved until deployment conditions change. Still, pairing Gemini-class reasoning with a commercial humanoid platform is a meaningful signal of where DeepMind sees the model going next — off the screen and into the world.