How does a single AI system teach humanoid robots to coordinate every movement from feet to fingertips while reasoning through hundreds of steps and collaborating with other machines?
Photo source:
Gemini-robotics
Most
robots today are pre-programmed for narrow, repetitive tasks. They execute
specific movements in a fixed sequence and cannot adapt when something
unexpected happens. Earlier AI models for robotics improved upper-body
manipulation but still left a critical gap. Robots could move their arms and
grippers, but they could not coordinate those movements with the rest of their
body. Walking, crouching, bending, and manipulating an object at the same time
remained beyond reach. Google DeepMind, the AI research division of Google,
released Gemini Robotics 2 to close that gap. It is a three-model AI robotics
model suite designed to give robots whole-body control, finer manipulation, and
the ability to work together on complex tasks.
The
system was introduced by Carolina Parada and the DeepMind robotics team. It
consists of three models: Gemini Robotics 2, a vision-language-action model
that processes visual inputs and instructions and converts them into full-body
motor commands; Gemini Robotics-ER 2, an upgraded embodied reasoning model that
crafts high-level task plans and enables natural-language communication between
humans and robots; and Gemini Robotics On-Device 2, the most efficient version,
optimized to run locally on robotic hardware without internet access. All three
are available in early-access form.
The
defining advancement in Gemini Robotics 2 is whole-body intelligence. Previous
versions focused primarily on upper-body tasks like picking up and placing
objects. Gemini Robotics 2 enables humanoid robots to coordinate movements
across their entire body simultaneously. In a demonstration on Apptronik's
Apollo 2 humanoid, the robot received a spoken instruction to put a watering
can into a green bin on a bottom shelf. It then walked across the room, picked
up the object, crouched down, and placed it on the correct shelf, all
autonomously without step-by-step commands.
The
system uses continuous video monitoring to track its own progress. By watching
live camera feeds, the robot detects when something goes wrong mid-task, adapts
its approach, and determines exactly when to move on to the next step.
According to Google DeepMind, Gemini Robotics 2 can automate tasks that
comprise hundreds of sequential steps, far beyond the short task sequences that
earlier models could handle.
A
key new capability is multi-robot collaboration. Gemini Robotics 2 enables
multiple autonomous machines to work together on a shared task. Rather than
operating independently, robots can divide responsibilities, coordinate their
movements, and complete complex work as a team. This is relevant for warehouse
logistics, manufacturing lines, and environments where a single robot cannot
handle an entire workflow alone.
Dexterity
has also improved. Testing showed stronger performance with two-finger grippers
than with multifinger hands, indicating that the AI can extract more precise
manipulation from simpler hardware. The On-Device 2 version can adapt to
entirely new two-arm robot designs with fewer than 200 training examples,
making it practical for rapid deployment on hardware the model was never
originally trained on. Apptronik has built a dedicated facility called Robot
Park where Apollo 2 humanoid robots collect real-world data used to train and
refine Gemini Robotics 2 models continuously.
Google
DeepMind introduced a new safety benchmark alongside Gemini Robotics 2 called
ASIMOV-Agentic. This benchmark tests whether robots can recognize unsafe
actions and refuse to carry them out. It also evaluates whether a robot knows
when to stop and request human assistance rather than proceeding with a task it
cannot safely complete. According to Google DeepMind, this represents a shift
from relying solely on physical safety cages toward embedding safety decisions
directly into the AI layer.
The
safety system combines conventional mechanical safeguards with controls built
into the AI models themselves. Google DeepMind has stated it is working with
experts, policymakers, and its own Responsibility and Safety Council to ensure
Gemini Robotics 2 operates safely alongside humans in shared environments. The
dual approach, mechanical barriers plus AI-level judgment, is intended to make
robots safer in unstructured spaces like homes, offices, and retail
environments where traditional safety cages are not practical.
Gemini
Robotics 2 arrives during an acceleration of investment and competition in
embodied AI. Multiple companies are building competing AI systems to serve as
the intelligence layer for humanoid and industrial robots. Google DeepMind's
approach treats robotics as another surface for the same Gemini model rather
than a separate specialized product. This means the general reasoning
capabilities that Gemini uses for text, images, and code are extended directly
into physical control, rather than building a robotics-specific AI from
scratch.
That
approach has strategic implications. If general-purpose AI reasoning transfers
effectively to whole-body robot coordination, it could reduce the advantage
held by standalone robotics AI startups whose value proposition depends on
having a robotics-specific model. The current partner list, which includes
Apptronik, Boston Dynamics, Agile Robots, and over 60 trusted testers such as
Agility Robotics, PAL Robotics, Universal Robots, and Rainbow Robotics,
suggests significant industry interest in testing whether this model delivers
on its capabilities outside of controlled demonstrations.
Please subscribe to have unlimited access to our innovations.