Apptronik's Apollo 2 runs Gemini Robotics 2 across three robot configurations
Google's Gemini Robotics 2 controlled Apollo 2 in three configurations, with reported dexterity results ranging from 32% to 92%.
By Ryan Merket · Published
Why it matters
Google gets real-world robot data and Apptronik gets a frontier model layer, but the disclosed task results show why commercial reliability remains the gating issue.
In a July 30 announcement, Google DeepMind introduced Gemini Robotics 2 and put the model in control of Apollo 2, the humanoid platform that Apptronik co-founders Jeff Cardenas and Nick Paine have spent a decade turning from University of Texas research into an industrial robot.
The release gives Cardenas and Paine a working version of the partnership they announced with Google DeepMind in December 2024: Apptronik supplies the body, deployment sites and physical training data, while Google supplies models intended to perceive instructions, plan tasks and control movement. In the accompanying X video, Google DeepMind cast Apollo 2 as the model's user, a playful treatment of a serious test for both companies.
Carolina Parada, the Google robotics leader who authored the Gemini Robotics 2 announcement, has worked across several of the perception problems now converging inside humanoid robots. She previously spent seven years leading speech work at Google and two years overseeing camera perception for self-driving vehicles at Nvidia. Her robotics groups at Google have focused on robot mobility and vision.
For Cardenas and Paine, Gemini Robotics 2 is a test of an old commercialization thesis. Cardenas worked in UT Austin's IC2 Institute helping researchers move technology out of laboratories and into markets. Paine earned three electrical and computer engineering degrees at UT Austin, developed the university's Series Elastic Actuator and worked on the NASA-Johnson Space Center team that built actuators and controls for the Valkyrie humanoid.
They co-founded Apptronik in 2016 out of UT Austin's Human Centered Robotics Lab. Apollo 2 packages that work into a robot platform designed for whole-body movement and dexterous manipulation. Its modular design supports bipedal and wheeled configurations, swappable batteries, fleet software and Apptronik's actuator technology. Gemini Robotics 2 is meant to make the hardware useful across tasks that cannot be reduced to a fixed sequence of programmed movements.
One checkpoint, several robot bodies
Gemini Robotics 2 is a family of three models. The main vision-language-action model converts visual and language inputs into motor commands. Gemini Robotics ER 2 handles higher-level reasoning, communication and planning across tasks that can last several minutes. Gemini Robotics On-Device 2 runs locally for settings where network latency or connectivity would be a liability.
Google says one Gemini Robotics 2 checkpoint controlled three configurations: Apollo 2 with SharpaWave hands, Apollo 2 with Inspire hands, and a Franka Duo robot fitted with a Robotiq gripper. That portability matters commercially. A model that must be rebuilt for every hand, sensor package or body shape would struggle to support a broad hardware market.
In Google's Apollo demonstration, a person asks the robot to put a watering can into a green bin on a bottom shelf. Apollo 2 walks to a table, retrieves the can, moves to the shelving and places it in the requested location. Other demonstrations show the robot tying knots, sealing a zip-top bag and manipulating a light bulb.
The SharpaWave hand has five fingers and 22 degrees of freedom, giving the model a much larger action space than a conventional two-finger gripper. Google also says its on-device model can be adapted to a new bi-arm robot using a few hours of data, typically fewer than 200 examples. Google offers a waitlist for early access to Gemini Robotics 2, while the Gemini Robotics model page links Gemini Robotics ER 2 to Google AI Studio.
The task results show where humanoids still fail
Google DeepMind's reported results put boundaries around the demonstrations. With Inspire hands, Apollo 2 recorded a 68.4% average success rate when picking objects from a table, 45.7% from the floor and 76.3% from a shelf.
Performance fell on several multi-finger tasks using SharpaWave hands. Google reported 36% success screwing in a light bulb, 44% tying a trash bag, 32% on a dustpan task and 40% sealing a zip-top bag. Unscrewing a bulb reached 92%.
Those are company-reported test results rather than independent measures of deployment readiness. They also show why dexterity remains a harder problem than producing a convincing video. A warehouse or factory customer will judge Apollo 2 through completed shifts, intervention rates, uptime, safety and cost per task. A robot that succeeds four times out of ten at sealing a bag still needs supervision or a narrower job definition.
Google says multi-finger dexterous manipulation remains challenging. The useful part of the release is that Google published failure-prone categories alongside stronger results. The range between 32% and 92% shows that dexterity is not one capability. Small changes in contact, grip and task geometry can produce very different outcomes.
Apptronik is building the data loop around Apollo
Apptronik unveiled Apollo 2 and an expanded, nearly 90,000-square-foot Robot Park in Austin on June 30. Apollo 2 robots there collect data through teleoperation and autonomous execution across logistics, manufacturing and retail tasks.
Cardenas described Robot Park as a continuous learning loop: robots perform work, collect data and feed those experiences into model development with Google DeepMind. Apptronik says similar data-collection workflows are operating with partners including Mercedes-Benz and GXO. Apptronik has not published enough operating data to establish production-scale reliability, so Robot Park should be understood as training infrastructure as much as a commercial deployment showcase.
Google gains access to a repeatable source of physical interaction data. Apptronik gains an intelligence supplier with the resources to train large multimodal models. That division lets Apptronik concentrate on actuators, hands, safety systems, manufacturing and fleet operations instead of financing a frontier-model effort alone.
Investors have funded the approach heavily. In February, Apptronik announced a $520 million Series A extension, taking the round above $935 million and total capital raised to nearly $1 billion, according to Apptronik. Participants included Google, B Capital, Mercedes-Benz, PEAK6, AT&T Ventures, John Deere and Qatar Investment Authority. Apptronik said the extension was priced at three times its earlier Series A valuation, though it did not state the absolute valuation in its announcement.
That capital is supposed to move Apollo from prototypes and data collection into repeatable production. The financing also makes the benchmark gaps harder to dismiss as routine research problems. Apptronik has the facilities, strategic partners and balance sheet to pursue commercial deployments. Reliability is the remaining test.
The hardware and model layers are splitting apart
The Google-Apptronik partnership reflects one route through the humanoid market. Figure AI is building a vertically integrated stack spanning robot hardware, models, tactile systems and manufacturing. Skild AI is pursuing a general-purpose robot-brain layer intended to work across different machines. Agility Robotics has concentrated on logistics work with its Digit robot.
Cardenas and Paine are placing Apptronik between those approaches. Apollo 2 is a proprietary hardware and operations platform, while its intelligence can draw on Google DeepMind's models. Apptronik remains responsible for the robot hardware, safety systems and fleet operations. Its Artemis system and Fleet Connect software cover perception, planning, controls, safety, task execution and management of deployed robots.
The result is a layered architecture rather than a robot controlled by a single model. Gemini Robotics can interpret goals and generate actions, while Apptronik's systems handle the physical machine and its operating constraints. That separation should make it easier to update intelligence without redesigning the robot. It also creates a dependency on how well two complex technical stacks work together outside a controlled demonstration.
Apollo 2 is both a product and a proving ground. Cardenas and Paine have built the body, the training facility and the customer relationships. Parada's team is supplying a model family that can walk that body across a room and manipulate objects with several kinds of hands. The disclosed results show meaningful progress, along with the distance between a broadly capable humanoid and a machine that can be trusted to finish every shift.