Robots that can only do one thing in one environment are already losing ground to systems that reason, adapt, and coordinate without being reprogrammed.
Pre-programmed robots are hitting a wall that DeepMind just decided to tear down
Most deployed robots run fixed sequences and fail the moment conditions change. Skills trained on one robot body do not transfer to another, and no single model has handled both high-level planning and low-level motor control at once.
Three models ship together and each one handles a different layer of the problem
Google DeepMind Ships Three Physical AI Models For Whole Body Control, Dexterity And Multi Robot Collaboration ships as three separate models: a vision-language-action model that converts visual and language input into motor commands for full humanoids and bi-arm robots, an embodied reasoning model built on Gemini 3.5 Flash that plans multi-step tasks over several minutes and communicates with humans, and a compact on-device VLA for edge deployment. One checkpoint drives the Apollo 2 humanoid with two different hand types plus a Franka Duo parallel gripper. You do not deploy all three to every robot; the stack is tiered by what the hardware actually needs.
Robotics engineers are the first to feel this, but they are not the only ones watching
- Robotics researchers who need a single model to generalize across multiple embodiments without retraining from scratch
- Automation engineers at manufacturers who are blocked by robots that cannot handle unpredictable part placement or collaborative handoffs between two arms
- AI safety teams who need a benchmarked foundation for physical AI risk, now that the ASIMOV-Agentic benchmark is live on Hugging Face under CC-BY-4.0
Physical AI is the category where the gap between lab demos and production deployment has been most brutal, and this release is the clearest signal yet that the gap is closing on a specific timeline.
Boston Dynamics and Figure have been racing toward this moment too
Competitors including Figure and 1X have each claimed general-purpose humanoid capability, but neither has shipped a public safety benchmark alongside a multi-model tiered stack in a single release. If dexterous manipulation scores, which currently range from 32 percent to 92 percent depending on the task, keep climbing at the rate the paper suggests, whole-body autonomous operation in unstructured environments moves from research preview to procurement conversation faster than most enterprise buyers expect.
What the stack lets you do right now
- Run whole-body humanoid control from feet to fingertips with one checkpoint
- Plan and execute multi-step physical tasks lasting several minutes using natural language
- Deploy a lightweight on-device VLA for robots without cloud dependency
- Evaluate physical AI safety posture using the open ASIMOV-Agentic benchmark dataset
Gemini Robotics ER 2 is available in public preview now; the VLA and on-device models remain in gated access.
Access tiers, not open weights
Pricing not listed — check our directory.
The five-finger dexterity scores tell you exactly where the ceiling still is
Multi-finger dexterity performance ranges from 32 percent to 92 percent depending on task complexity, which means fine manipulation in unstructured settings is still unreliable enough to block production deployment for precision-sensitive use cases.
The rest of the field is not standing still
Physical Intelligence’s pi0 model targets a similar whole-body control problem from a diffusion policy angle rather than a VLA-plus-reasoning split. If you need on-device deployment without Google’s access restrictions, OpenVLA from Stanford remains the most documented open-weights alternative at smaller scale.
The race to put reasoning inside the robot body just entered a new phase
This is one of the most consequential infrastructure shifts in physical AI in 2025, and the downstream effects on automation procurement, safety standards, and robot hardware decisions will move fast. We cover tools like this every Friday — subscribe here and we’ll send the best ones straight to you.