§18.4 Reasoning + action: where LLM-style chains of thought meet control

What “reasoning before acting” actually means here

The idea has a concrete form, and it is worth pinning down before the tradeoffs make sense. An embodied chain-of-thought policy, of which ECoT (Zawalski et al., 2024) is the canonical early version, does not go straight from image-and-instruction to action tokens. It first generates intermediate text: a description of the scene, the sub-task it should tackle, sometimes the coordinates of the relevant object or a plan in words, and only then the action. The reasoning is generated by the same model, in the same autoregressive stream, so the action is conditioned on the model’s own written-out plan rather than on the raw inputs alone. §8.2’s Decision Transformer showed that control can be sequence modeling; embodied CoT extends the sequence backward, so that language reasoning tokens precede the action tokens the robot executes.

The reported payoff matches the language-model intuition. Reasoning helps most on the tasks that need it: long-horizon tasks where the plan matters (§18.2), tasks with linguistic ambiguity where the model has to interpret “the cup next to the blue one,” and out-of-distribution situations where a policy benefits from thinking rather than reflexively pattern-matching. ECoT and its successors show generalization gains on exactly these, which makes sense, because the reasoning is where a novel instruction gets decomposed into something the low-level action head has seen before.

It also matches the intuition in the opposite direction, and this is the part people forget: reasoning does not help on tasks that do not need it, and it can hurt. A short, unambiguous grasp the policy has done ten thousand times gains nothing from a paragraph of deliberation, and the deliberation adds latency and a fresh chance to talk itself into the wrong plan. This is the same lesson §14.1 taught about dual systems, that not every action needs the slow path, and it means embodied reasoning is not a universal upgrade you bolt onto every policy. It is a tool for the hard fraction of tasks, and knowing which fraction is itself a decision the system has to make, either by a fixed schedule or, harder, by the model recognizing when it is confused enough to be worth thinking. That triaging problem, when to reason at all, is less studied than the reasoning itself and is a clean place for new work.

The cost, and why it is not a footnote

There is a hard problem baked into this, and it is not incidental: reasoning is slow. Generating a paragraph of chain-of-thought before every action means running the language model to produce dozens or hundreds of tokens, and §14.4 spent a whole section on why a VLA that cannot close its control loop fast enough is not deployable. A policy that stops to write an essay before each motion cannot run at 20 Hz. This is the tension the whole subfield is organized around, and it is the same tension §14 resolved with dual systems: a slow reasoner and a fast controller running at different rates, the reasoner setting direction a few times a second and the controller producing smooth motion in between. Reasoning-plus-action, in its deployable form, is often a dual-system architecture wearing the reasoning label, with the chain of thought living on the slow clock.

Gemini Robotics-ER (arXiv:2510.03342), introduced in §14.3, is the industrial version of this bet. DeepMind’s “Embodied Thinking” is embodied chain-of-thought at foundation-model scale, a model that reasons about the physical scene, spatial relationships, object affordances, the steps a task needs, and then drives action. It is the third or fourth entry in the dual-system family the book has been tracking, and its distinguishing feature is that the high-level system’s deliberation is explicitly reasoning rather than a black-box latent. That choice is the §17.5 analyzability argument made into a product decision: a system whose intermediate reasoning is legible text is one whose mistakes you can sometimes read, which matters for a technology nobody can formally verify.

The clever fix: reason in training, not at inference

The most interesting recent move sidesteps the latency problem outright, and it is worth understanding because it is the kind of idea a student could have and extend. ERVLA (arXiv:2606.03784) asks a sharp question: if chain-of-thought helps the model learn, does the model still need to generate it at test time? Their answer is no. ERVLA trains with reasoning traces, so the model learns from the decomposed, spelled-out plans, but uses a CoT-dropout strategy so that at inference it predicts actions directly without generating the reasoning tokens. The reasoning shapes what the model learns and then gets out of the way when the clock is running. The reported numbers are strong, roughly 86.9% on LIBERO-Plus and 53.2% on VLABench, and the mechanism is the point: you can get the generalization benefit of reasoning as a training signal while paying none of the inference latency, because the essay was a scaffold, not a runtime requirement.

This resolves, at least partially, the tension the rest of the section built up. If reasoning’s value is that it organizes learning, then it belongs in the loss, not in the forward pass, the same way a good curriculum shapes a student who then answers fluently without narrating every step. The open question ERVLA opens rather than closes is how much reasoning has to survive to inference. Some tasks may genuinely need runtime deliberation, a truly novel situation the training traces never covered, and dropping the chain of thought entirely would give up the ability to think on the fly that motivated the whole idea. Where the line sits, between reasoning-as-training-signal and reasoning-as-runtime-capability, is unsettled and is one of the more tractable open problems in the chapter.

There is a subtler worry the subfield has started to confront, and it is worth stating because it cuts against the whole premise. When a language model writes a chain of thought, we assume the reasoning caused the answer. That assumption is shakier than it looks: a model can produce plausible reasoning that has little to do with how it actually arrived at the action, a rationalization rather than a computation. For a robot this matters more than for a chatbot, because §17.5’s analyzability argument rests on the reasoning being faithful, on the written plan actually reflecting what the policy will do. If the chain of thought is a story the model tells that does not bind its motor output, then legible reasoning buys the appearance of transparency without the substance, and a monitor reading the reasoning would be reassured by a plan the robot was never going to follow. Whether embodied chain-of-thought is faithful in this sense, whether the reasoning constrains the action or merely decorates it, is open and under-measured, and it is a place where a careful experiment, perturb the reasoning and see if the action changes, would say something the field needs to hear.

A related thread: pointing, not just planning

Reasoning about a scene is not only about task decomposition; it is also about grounding language in specific pixels, and a parallel line of work treats that as the reasoning to prioritize. Embodied-R1 (arXiv:2508.13998) and related systems emphasize spatial reasoning, getting the model to localize the object an instruction refers to, output a point or a region, and act on that grounding, which is the failure mode where a policy understands “pick up the wrench” as language but reaches for the wrong tool. This is reasoning in a narrower, more checkable sense than a full chain of thought, and it has the property that its output is verifiable: a predicted grasp point either lands on the wrench or does not, where a paragraph of reasoning is harder to score. For a reader deciding where to contribute, grounded spatial reasoning has the advantage that you can measure whether it worked.

Reasoning joined to action stopped being a thought experiment during this book’s writing and became a crowded, fast-moving subfield with a clear internal logic. The promise is real, deliberation helps on the long, ambiguous, novel tasks that reflexes fail. The cost is real too, reasoning is slow and control is fast, and the field’s best answers are to split the clocks the way §14 did or, more cleverly, to use reasoning as a training signal that never has to run at inference, which is ERVLA’s bet. What none of it has yet delivered is a robot that reasons its way through a genuinely new physical problem in real time, the way the language-model analogy promises; that remains ahead. With the four research bets of this chapter now on the table, cross-embodiment transfer, long-horizon dexterity, video pretraining, and reasoning, the next section steps back to ask what a reader should read next, and where a person with one robot and a year can actually push.