“When a humanoid composes its next move as a sentence of small decisions, it stops acting like a puppet and starts behaving like a mind in motion. Autoregressive control is how we teach bodies to think about what they are about to do.”

— Aditya Mohan, Founder, CEO & Philosopher-Scientist, Robometrics® Machines

In our earlier piece, When Actions Think Autoregressive Heads and Chains of Thought, we treated the autoregressive action head as a kind of grammar for decisions. Instead of throwing a single vector at the world, the policy composes a short sequence of choices, each conditioned on the last. In robotics, and particularly in humanoid control, that vantage point becomes far more than a metaphor. A humanoid is not a joystick with legs; it is a dense forest of joints, contacts, and constraints. To move it convincingly through a messy world, we need actions that are spoken in small, coherent phrases rather than shouted as one undifferentiated command.

Formally, an autoregressive control policy still implements a probability distribution over actions, but it factors that distribution into ordered pieces: p(a) = ∏ⁿᵢ₌₁ p(aᵢ ∣ a₁, …, aᵢ₋₁). Each aᵢ is a different aspect of the motor decision. One head selects a high‑level skill token—step, reach, grasp, shift weight. Another refines timing and duration, deciding when to begin and how long to commit. Pointer‑style heads choose which limbs participate, treating effectors (hands, feet, torso segments) as items in a list. A spatial decoder, often a deconvolutional network operating on egocentric depth or occupancy grids, paints a heatmap over reachable space and samples specific footholds or grasp points. In a modern Transformer‑based policy, these heads sit on the same latent state, embedding each sampled choice and feeding it forward so the next decision is made with full awareness of what came before. In more concrete terms, a pointer network is just an attention layer whose output is a probability distribution over input indices rather than over a fixed label set, and the spatial decoder is a CNN run "in reverse" (a deconvolutional decoder), upsampling compact visual features into dense grids of action probabilities that align with pixels or map cells.

If you are not steeped in sequence‑model jargon, three notions help fix ideas. A causal Transformer is a neural architecture that reads a sequence left to right and uses attention to let each position look back at earlier positions, but never forward in time. An egocentric depth map or occupancy grid is simply a sensor‑derived picture of free space and obstacles in front of the robot, expressed from the robot’s own point of view. Teacher‑forcing, which we will mention later, is a standard training trick where we feed the model the ground‑truth next action during learning instead of its own prediction, so it can learn the correct continuation before being asked to roll out actions autonomously.

Consider a humanoid crossing a cluttered workshop. At every control step it must decide whether to advance, sidestep, pause, or reach out to steady itself on a beam. A flat policy would emit a long vector of target joint angles or torques in one breath, implicitly solving skill choice, timing, limb selection, and contact placement in a single opaque draw. An autoregressive head does this in articulated stages: first it chooses a locomotion primitive (for instance, a lateral step), then a phase offset to keep its gait consistent with the previous stride, then the participating limb (left or right foot, or both for a hop), and finally the exact contact patch on the floor, drawn from a dense probability map that already knows about spilled tools and uneven planks. The resulting sequence is not slower in physical time—the whole chain unfolds within a single policy call—but it is far more structured in the policy’s internal time.

This structure pays off because humanoid actions are combinatorial. Choices that should be tightly coupled often live in different slices of the control space: which hand you use to catch a falling object depends on where the object is, how the torso is oriented, and what the feet are doing to keep balance. Autoregressive heads let us express those couplings explicitly. Instead of enumerating every possible joint configuration, the policy traverses a compact decision tree, leaning on specialized heads that know how to operate over sets (pointer networks), over continuous manifolds (spatial decoders), and over discrete skill vocabularies. Work on autoregressive policies for continuous control and manipulation has shown that this factorization improves exploration efficiency and stability in both simulated and real robotic domains, compared to monolithic action distributions or naïve discretizations.

The natural worry is latency: if language models already pay a cost for Chain‑of‑Thought reasoning, do autoregressive actions slow a robot at the exact moment it must be decisive? In practice, recent work on robotic manipulation suggests the opposite. The Chunking Causal Transformer (CCT) is a variant of the standard causal Transformer that predicts not just the next single token, but a small block of future tokens in one forward pass: past actions occupy the left part of the sequence, a block of “empty” future positions sits on the right, and CCT fills those empty slots jointly while still respecting causal ordering. That means the model remains autoregressive in spirit—future actions depend on past ones—but it needs far fewer passes through the network to produce the same horizon of control, which keeps inference fast even at high control frequencies.

Built on top of CCT, the Autoregressive Policy (ARP) wraps this chunked Transformer into a full control stack: it ingests past actions and current visual features, then emits a short chunk of structured future actions that may include discrete skill identifiers, continuous waypoints, gripper commands, or contact flags in the same sequence. During training, ARP uses teacher‑forcing: ground‑truth action chunks from demonstrations are fed in as targets, and an attention‑interleaving pattern alternates between self‑attention over the action tokens and cross‑attention into the vision features. The result is a policy that can learn long, hybrid action sequences, yet at run time only needs a handful of chunked autoregressive steps per update. Experiments across diverse benchmark suites show that such autoregressive policies match or beat environment‑specific architectures while using fewer parameters and less compute, which is precisely what we want for humanoids operating under tight latency and power budgets. Unlike textual Chain‑of‑Thought, which might spill dozens of tokens, an autoregressive action head for control might only roll out four to seven micro‑decisions before handing the result to the motors.

On accuracy and complexity of behavior, the evidence is even stronger. Recent dense autoregressive policies that expand sparse keyframes into full action trajectories have shown that modeling actions as sequences, rather than single holistic predictions, yields richer, more precise motion across many tasks. For a humanoid, that translates to a single architecture that can walk, climb a ladder, reach for a tool, or brace against a shove without swapping controllers. The same sequence model that once predicted the next word in a sentence is now predicting the next sliver of a whole‑body movement, capturing causal dependencies between limbs and contacts in a way that flat vectors never quite manage. Autoregressive control is not the only path to intelligent movement—classical motion planners, model‑predictive control, and handcrafted state machines all hold vital lessons—but it is rapidly becoming one of the most powerful ways to fuse perception, decision, and actuation into a single, trainable system.