# Apptronik's Apollo 2 runs Gemini Robotics 2 in dexterity demo > Source: > Published: 2026-07-30 19:20:00+00:00 Jeff Cardenas' Apptronik put its Apollo 2 humanoid at the center of [Google DeepMind's Gemini Robotics 2](https://deepmind.google/models/gemini-robotics/vla/?ref=runtimewire) launch on July 30, with a [demo on X](https://x.com/GoogleDeepMind/status/2082904278461469091?ref=runtimewire) showing the robot working through fine-motor tasks and kitchen chores. Google DeepMind says the vision-language-action model can control a humanoid from its feet to its fingertips, pushing Gemini Robotics beyond tabletop and arm-only systems. [Google DeepMind on X](https://x.com/GoogleDeepMind/status/2082904278461469091?ref=runtimewire) For [Jeff Cardenas](https://apptronik.com/company/leadership?ref=runtimewire), Apptronik's co-founder and CEO, the launch is evidence for a strategy he has pursued since spinning Apptronik out of the University of Texas at Austin in 2016. Cardenas, who studied business and technology commercialization at UT Austin after working at Deloitte and the university's IC2 Institute, built Apptronik around turning robotics research into products. Apptronik has paired its hardware and data operation with Google DeepMind's robotics models. That division of labor is visible in the demo. [Apollo 2](https://apptronik.com/apollo/apollo-2?ref=runtimewire) supplies the actuators, hands, sensors and mobile body. Gemini Robotics 2 converts images and language into motor commands, coordinating balance and manipulation as Apollo 2 bends, reaches and handles objects. The video attachment is labeled "Autonomous 1x," and Google DeepMind said the demo pushed Apollo 2 to complete tasks requiring the dexterity people rarely have to think about. ### The scores expose the remaining gap Google DeepMind's own evaluations put useful numbers around the polished footage. On general whole-body pickup tasks, an Apollo robot achieved a 68.4% success rate picking up objects from a table, 45.7% from the floor and 76.3% from a shelf, according to the [Gemini Robotics 2 model page](https://deepmind.google/models/gemini-robotics/vla/?ref=runtimewire). Performance varied on multi-finger tasks using Sharpa hands. Google DeepMind reports 92% success unscrewing a light bulb, 40% closing a Ziplock bag and 44% tying a trash bag. Those results are research evaluations. They do not prove Apollo 2 can reliably run a kitchen or work an unattended shift. They also show why dexterity has become a central problem for humanoid developers. A robot can recognize a bag, plan to close it and move its hand toward the seal, yet still fail because the plastic bends, the fingers miss their grip or the robot applies force at the wrong angle. Google DeepMind is making a broad claim about portability across hardware. Gemini Robotics 2 is designed for tabletop robots, dual-arm systems and full humanoids, including systems with complex hands. The direct-control model remains in private preview. Google DeepMind says the Gemini Robotics 2 VLA and On-Device models are available to early-access partners through its [Trusted Tester Program](https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/?ref=runtimewire). The release also includes [Gemini Robotics ER 2](https://deepmind.google/models/gemini-robotics/embodied-reasoning/?ref=runtimewire), a higher-level model for spatial reasoning and multi-step planning. A third model, [Gemini Robotics On-Device 2](https://deepmind.google/models/gemini-robotics/on-device/?ref=runtimewire), is designed to run locally when network latency or connectivity would make cloud inference impractical. Google DeepMind says On-Device 2 can adapt to a new robot with fewer than 200 examples and several hours of training. It is available to early-access partners through the Trusted Tester Program. ### Apptronik built a data operation around Apollo The launch follows nearly two years of joint work. Apptronik and Google DeepMind [announced their partnership](https://apptronik.com/news-collection/apptronik-partners-with-google-deepmind-robotics?ref=runtimewire) on December 19, 2024, joining Gemini-based robotics models with hardware descended from work at UT Austin's Human Centered Robotics Lab. Cardenas' co-founders supplied much of that technical lineage. [Nick Paine](https://apptronik.com/company/leadership?ref=runtimewire), Apptronik's [chief technology officer](/article/rob-witoff-joins-coinbase-as-cto), completed his electrical and computer engineering degrees at UT Austin, developed the UT Series Elastic Actuator during his graduate work and helped build actuator controls for NASA's Valkyrie humanoid. Luis Sentis, a UT Austin professor and Apptronik adviser, led the lab from which Apptronik emerged. Cardenas has framed the goal as building a single robot capable of learning thousands of jobs. Apptronik's commercial focus starts with manufacturing and logistics, where repetitive picking, kitting and material-handling tasks offer more structured environments than a home kitchen. The kitchen demo matters because bags, tools and loosely placed objects test whether the same system can handle variation instead of repeating a fixed industrial motion. Apptronik opened an expanded [Robot Park](https://apptronik.com/news-collection/welcome-to-robot-park-where-apptroniks-apollo-goes-to-work?ref=runtimewire) in Austin on June 30 alongside the public unveiling of Apollo 2. The nearly 90,000-square-foot facility runs Apollo 2 robots in bipedal and wheeled configurations, gathering data through autonomous operation, teleoperation and simulation. Apptronik says similar collection systems operate at Google DeepMind and customer sites including Mercedes-Benz and GXO. That continuous data loop is the commercial bet underneath the July 30 demo. Apptronik supplies physical experience from robots performing real tasks. Google DeepMind uses that experience to train and refine models that can return to the same hardware with broader capabilities. Each side gains something costly to build independently: Apptronik gets an advanced AI partner, while Google DeepMind gets access to humanoid hardware and task data. ### Investors have financed the hardware side of the bet Building that loop has required an unusually large amount of early-stage capital. On February 11, Apptronik [announced a $520 million Series A-X](https://apptronik.com/news-collection/apptronik-closes-over-935-million-series-a?ref=runtimewire), bringing its Series A financing above $935 million and total capital raised to nearly $1 billion, according to Apptronik. Existing backers included B Capital, Google, Mercedes-Benz and PEAK6; AT&T Ventures, John Deere and Qatar Investment Authority joined the extension. In its [financing announcement](https://apptronik.com/news-collection/apptronik-closes-over-935-million-series-a?ref=runtimewire), Apptronik said the extension was opened at three times the original Series A valuation. Apptronik did not state a dollar valuation in its announcement, while [TechCrunch reported](https://techcrunch.com/2026/02/11/humanoid-robot-startup-apptronik-has-now-raised-935m-at-a-5b-valuation/?ref=runtimewire) a post-money valuation of about $5.3 billion. The capital gives Cardenas room to fund hardware production, training facilities and long deployment cycles before humanoids reach dependable commercial operation. It also puts pressure on Apptronik to turn demonstrations into repeatable work. Rival Figure AI has chosen to build its Helix intelligence stack in-house, while Apptronik is tying Apollo's progress closely to Google DeepMind's model roadmap. Gemini Robotics 2 strengthens that partnership by giving Apollo 2 a more capable control layer and by publishing enough evaluation data to show where the system still breaks. The strongest result, unscrewing a bulb at 92%, sits beside basic manipulation tasks that succeed less than half the time. Cardenas' bet will be judged by how quickly Apptronik and Google DeepMind can close that reliability gap across long shifts, changing objects and customer sites.