What is embodied reasoning? The term, its senses, and the Gemini ER era

Embodied reasoning is a robot's skill at tying what it sees, what it plans, and what it does into one loop. The plan then answers to the real world, not just to a text prompt. But the phrase has more than one owner. This page splits the four senses apart. It names who uses the phrase, who avoids it, and what to ask the next person who says it. Every check on this page is dated October 26 2026.

The short answer first. In robotics, embodied reasoning is thinking tied to a body in a place. The model reads cameras and joint state. It weighs space, physics, and progress, then picks what should happen next. Google states the idea in its own docs. Its Gemini Robotics ER family is a set of vision-language models that let robots sense the physical world, read visual data, reason about space and time, plan steps, and drive tools (Google robotics docs). But the term is not a settled standard. Google brands a product line with it. Research papers use it as a skill name. NVIDIA's own glossary has no entry for it. The largest robotics glossary online carries none either. Both were checked on October 26 2026. A page about embodied reasoning that never says which sense it means has already failed the reader. Each section below says which sense it means.

Last verified: 2026-10-26

Embodied reasoning

The core sense is a skill, not a product. A robot with this skill links three things that text models keep apart. One is sensing: camera frames, depth, joint angles, and video over time. Another is thinking: plans, spatial calls, and guesses about what happens if it acts. A third is action: the commands that move the arm or the base. The word embodied carries the load. Thought here sits inside a real scene. It has a body, a place, gravity, and clutter. A chat model can talk about grasping. An embodied reasoner must answer for a grasp. Does the gripper fit? Is the object heavy? Did the last try fail?

The research field states this plainly. A CVPR 2026 workshop named Embodied Reasoning in Action opens with a definition. Embodied reasoning, it says, is the skill of grounding what is seen, said, and done in the physical world, and it is now central to robot manipulation (ERA workshop). The workshop runs two public tests of the skill. RoboSpatial holds 350 spatial question pairs from five real homes. PointArena holds 982 image pairs that test precise pointing in plain language. The site notes that frontier models, such as Qwen3-VL and Gemini Robotics, now use these tests to grade their own spatial skill. That is the field sense of the term: a real skill, with scoreboards.

Two lines keep the meaning honest. First, embodied reasoning is not motor control. The fast loops that drive joints run at hundreds of cycles per second. That layer is different, and it never leaves the robot. Second, embodied reasoning is not the action policy itself. Google calls its ER 2 model a high-level brain for robots. It hands motor work to a lower level vision-language-action model, a VLA, which you supply (DeepMind model page). The reasoner looks, thinks, and hands off. Motion belongs to the action model. If a seller offers "embodied reasoning" and you expect arm motion in the box, that is a unit error on your side, not a fixable bug on theirs.

Embodied reasoning model

The second sense is a model class. This is the Gemini ER era of the title. Through 2025 and 2026, the phrase moved from paper titles into product names. Google now sells a line whose model names carry the term as their middle letters. Google's own pricing page states the link. Gemini Robotics ER 2, it says, is short for Gemini Robotics Embodied Reasoning 2. The page calls it a vision-language model endpoint that helps robots read their surroundings with precision (Google pricing page). That family meaning sits in the developer docs. Gemini Robotics ER, embodied reasoning, models are vision-language models that let robots sense and act in the physical world (robotics docs).

The line has a dated past, and the dates matter, because tutorials go stale. ER 1.5 started the family in 2025. It was the first model in the line open to all developers. ER 1.6 landed on April 14 2026. That version added pointing, success checks, and gauge reading. Google then retired ER 1.6 on August 31 2026. The docs now say to move to ER 2. ER 2 is the current row, in public preview on Google AI Studio and the Gemini API. The full family tree with dates is on our Gemini rate card guide. Search still ranks ER 1.6 explainers today. Check the date on anything you read about this family.

Google is not alone in putting the words inside a model name. The Allen Institute for AI ships a model whose card opens this way. Molmo2-ER, Embodied Reasoning, is a 4B vision-language model built for the sensing skills that action models depend on (model card). Its maker reports that it beats every open-weight baseline, plus Gemini Robot-ER 1.5 Thinking and GPT-5, on 9 of 13 embodied reasoning tests. A group at Zhejiang University ships Embodied-Reasoner. They call it an embodied reasoning model for the physical world. It plans long tasks, tracks scene state, and reflects on its own actions (repository). So the model-class sense has two flavors. One is a single company's product line. The other is a wider family of open models named for the skill. When you hear "embodied reasoning model", ask which flavor.

Embodied chain-of-thought

The third sense is a training method, and it is older than the product names. In 2022, a Google team put the phrase in a paper title: Inner Monologue: Embodied Reasoning through Planning with Language Models (Google Research). The idea was simple. A language model could think through a robot task as a quiet inner talk. It would fold in feedback from success checks, object checks, and humans. When the world disagreed, it would replan. That paper is why older forum threads use this phrase with no product in sight.

The method grew into embodied chain-of-thought, or ECoT. Work from Berkeley, Stanford, and Warsaw trains a vision-language-action policy to write out its steps before it acts. The steps name the task, the plan, the sub-task, the motion, the gripper spot, and boxes around the objects that matter (ECoT project page). Grounding is the point. Plain chain-of-thought talks about sub-tasks in words. The embodied kind anchors its words in the camera frame and the robot state. On their new-task tests, the team reports a 28 percent absolute gain in success rate for OpenVLA, with no extra robot data. A later line, Embodied-R1, trains a 3B vision-language model for embodied reasoning and pointing. Its paper reports 56 percent zero-shot success in a standard test world, and 87 percent across 8 real arm tasks (arXiv). Those are the makers' own numbers, quoted as such.

This sense explains a trap you will meet in test results. Some tests grade the reasoning trace itself. Others grade task success, where the thinking is only a means. When a scoreboard says "embodied reasoning", find out which of the two it counts. A model can explain well and still fail the grasp.

Embodied cognition

The fourth sense is the oldest, and it is not about robots. Embodied cognition is a claim about human minds from psychology and philosophy. The Wikipedia article on it is a psychology article. Robotics appears in it only as one subsection under applications, and the robot sense of the term has no article of its own, checked October 26 2026. The article describes a set of theories on how thought is shaped by the body and its skills (Wikipedia). In that view, thinking is not a brain acting alone. It is a body at work. The literature cites gestures while talking, pacing while puzzling, and the feel that a hard problem is "heavy" as thought leaning on the body.

The robot field borrowed the word from here, and the debt shows up in search. Ask for the meaning of embodied reasoning. The first page mixes robotics rows with dictionary entries for the word embodied, a philosophy magazine, and psychology essays. People-also-ask boxes mix the two crowds in one list. Questions about robot models sit next to questions like, what does embodied mean in psychology. Neither crowd is wrong. They are asking about two different things that share two words.

The rule is simple. If the source talks about human minds, lab studies, or philosophy, it means the psychology claim. It will not help you pick a robot model. If the source talks about models, cameras, tests, or endpoints, it means the robot skill. This page is about the robot senses, and it says so.

Embodied AI

The neighbor term is embodied AI, and knowing it saves you a search. NVIDIA's glossary defines embodied AI as AI built into physical systems that then act in the real world. Its examples are robots, humanoids, vehicles, and warehouses (NVIDIA glossary). NVIDIA also brands the field Physical AI. Yet the same company's research titles use the phrase embodied reasoning. Its Cosmos-Reason1 paper is called From Physical Common Sense To Embodied Reasoning. The paper builds a frame for reasoning that spans different bodies (arXiv). One vendor can use the field name in its glossary and the skill name in its papers.

Here is the proof that the skill term is still young. NVIDIA's glossary has no entry for embodied reasoning. Its page address returns not-found, checked October 26 2026. The largest robotics glossary online, almond.bot, lists 1,052 terms. It has Embodied AI, Embodiment, and Embodiment Gap, but no Embodied Reasoning entry, checked the same day. Wikipedia points the concept at the psychology article. So the field has a name everyone uses. The skill has a name the research world uses. One vendor has put the skill name in a product line. That is what the Gemini ER era of the term means. The letters ER now do double duty. They name a Google product class, and they initial a research field.

Who uses the term, and who does not

Every row below was checked on October 26 2026. Read the last column as a fact with a date, because glossaries move.

One more demand fact, checked the same day. Google's own search suggest for embodied reasoning returns ten rows. One of them is "embodied reasoning er model". The product name has fused into the plain term deeply enough that readers now type the collision itself. That row is why a term page like this one needs to exist.

What to ask when someone uses the term

Worked example. A vendor page says their robot runs embodied reasoning. You have four questions, in order.

1. Which sense? Skill, model class, training method, or a borrowed psychology frame? If they cannot say, the words may be decoration.
2. Where does action happen? A reasoner hands off motion. Ask what drives the motors. Their VLA, your VLA, or nothing yet? The handoff is the whole design.
3. What is graded? Ask for the test name and the date. RoboSpatial and PointArena are public. A vendor score with no test name is a claim, not a result.
4. What do you pay, in which unit, on what date? Tokens, calls, or actions? The units differ across the market. The pricing units guide converts them.

If the answers are a named sense, a named VLA, a named test, and a dated rate in a stated unit, the phrase is being used honestly. Anything less, and you are reading marketing.

Hosted embodied reasoning, dated

Who serves an embodied reasoning model today, on a public route? Checked October 26 2026.

Three honesty notes on that table. First, the two expert hosts are real and new. They serve the action tier, not the reasoning tier. That split is the market's current shape. The wider census of which big hosts serve which robot models has its own guide. Second, Google's card is the reference price for the class, and its date behavior is odd in a useful way. The doubling is printed in each cell, not buried in terms. A line-by-line read of that card, with its limits and request shapes, is the Gemini rate card guide. Third, the empty rows carry dates, because empty rows expire. Any web page saying nobody will ever host a model class is making a guess, not stating a fact.

This platform's own ER-class row, in the fence, with its basis:

The model's own page holds the use cases and call shapes: molmo2-er. Its cost story, with the token math worked out and set against Google's card, is the Molmo2-ER pricing guide. Live rates for every listed row render on the pricing page. The action tier that a reasoner hands off to is explained in the VLA models guide. And the program that serves policies over a wire is defined in the policy server guide.

Frequently asked questions

Is embodied reasoning the same as embodied AI?

No. Embodied AI is the field. It means AI in physical systems of any kind, and it is the umbrella term in NVIDIA's glossary. Embodied reasoning is one skill inside that field. It ties sensing, planning, and action to one physical scene. Every embodied reasoner is embodied AI. Most of embodied AI is not embodied reasoning. Action policies and sensing tools belong to the field, not to this skill.

What does embodied mean in psychology?

It means thought shaped by the body. In psychology, embodied cognition is the claim that thinking leans on body state, movement, and setting, not on symbols alone. The robot field took the word from this literature. Search results for the phrase still mix the two crowds. Check whether a source is about human minds or robot models before you use it.

Is Gemini Robotics ER a VLA model?

No. Google's own pages call ER 2 a high-level brain. It senses, plans, tracks success, and calls tools. Then it hands motor work to a lower level vision-language-action model that you supply. It reasons about action and does not emit motor commands itself. The action side of that split is covered in the VLA models guide.

Where did the term come from?

The words have lived in psychology for decades. In robotics, a 2022 Google paper put them in a title: Inner Monologue: Embodied Reasoning through Planning with Language Models. Test suites and chain-of-thought training kept the phrase alive in research. Google's ER product line, which names the term inside its model names, made it a buying keyword in 2025 and 2026.

Does anyone host an embodied reasoning model today?

Yes, two kinds, dated October 26 2026. Google serves its ER 2 in public preview, metered per million tokens. This platform lists molmo2-er, an open ER-class model from the Allen Institute for AI, on its chat route. It has been listed since October 6 2026. Its rate sits in the fence above and on the pricing page. The expert robot hosts serve action models, not ER-class models, as of the same date. To try the platform's row, sign up and create a key.

The term is young enough that its own meaning is contested ground. One vendor's product name, one field's skill name, one discipline's older claim, and one training method all answer to it. Name the sense, name the owner, date the claim. The model-class story continues on the Gemini rate card guide and the Molmo2-ER pricing guide. Next door, the action tier is the VLA models guide. Rates for every listed row render on the pricing page.

Last verified: 2026-10-26