Research

A dancing robot still needs a person to decide

On 13 November 2026 twelve researchers published a proposal for what they call agentic robotics: a robot built in three layers, a way to test the whole system, and four levels of autonomy. We placed Paqarina inside it. Our robot has the two lower layers, and the layer that decides is still a person with a remote control.

The proposal in a few lines

The paper is called Envisioning Agentic Robotics: A New Era of Robotic Systems.

It starts from a difference. In the older picture a robot is one learned policy: an instruction goes in and movement comes out. In the new picture a robot is an agent: it pursues a goal by choosing among skills and tools, checking what happened, and changing its decision.

Then it describes three layers.

LayerWhat it doesWhat the paper puts there
System 2Decides. It reads what the person wants, chooses skills and tools, checks the result and changes the planA large model, and around it a harness: a planner, a memory, a verifier, an interface to tools and a manager of context
System 1Holds the skills that can be calledLearned movement skills, perception, motion planning
System 0Moves the body in real timeWhole-body tracking, and the control of each joint

In the paper the word harness means the software that surrounds the model. It is not the strap that holds a robot up.

Beside the three layers runs a small monitor. It watches the person, the body and the scene, and it can stop the robot.

Where Paqarina sits

LayerWhat Paqarina hasWhere it runs
System 0The network that follows a dance and keeps the balance: 29 angles, 50 times a second. Under it, the motors' own controlOn the robot
System 1Eight dances. Each one is a skill, started from a remote controlOn the robot
System 2A person, who chooses the dance, starts it, watches, and decides what to do after a fallBeside the robot
The monitorA person standing next to the robotBeside the robot

The paper says something close about the robots that work today: where a robot is trusted with more, a person in the background is doing the work of System 2.

Violeta and the robot. For now, the layer that decides is a person. Photo: Paqarina.
Violeta and the robot. For now, the layer that decides is a person. Photo: Paqarina.

The parts of a System 2 that already exist

Paqarina has built parts of what the paper calls the harness. They work, and none of them runs on the robot. They run on a laptop, days before.

PartWhat Paqarina hasWhen it runs
MemoryThe archive: 137 takes of our own, kept as 26 points for every instantBefore
PlannerA score in words becomes a remix that respects a stage and a partnerBefore
VerifierEvery dance runs 20 times in a second simulator. A planned sequence is checked against the robot's jointsBefore
Interface to toolsThe chain itself, from four cameras to 29 angles. A person runs each stepBefore
Safety gateA hoist and a personDuring

So the distance between Paqarina and an agentic robot is not that the parts are missing. It is where they are and when they run. Ours decide before the dance. An agent decides during it.

Four levels, measured on a dance

LevelIn the paperFor a robot that dancesPaqarina today
1Carry out one skillDance one dance from beginning to endIn simulation, 8 dances of 8. On the real robot, 6 seconds of one
2Complete a whole task without helpReceive a score and a stage, and dance it: choose the pieces, stay inside the circle, begin and end standingThe choosing is done ahead of time, on a laptop. The robot does not choose
3Coordinate steps that depend on each otherHold a place in a piece with dancers and avatars: an entrance, a partner at a distance, the music as the clockWritten by hand in the app, and planned as a duo. Not on a robot
4Hold a role over a long timeBe a member of a company: rehearse, learn a new dance, remember a stageNot started

On this ladder the simulated robot stands on the first step, and the real one is below it.

Test the system, not only the model

For us the strongest point of the paper is about testing. It says that how well a model decides, and how well the whole system works with the world running, have to be measured separately. It also says the result should be reported as a profile across nine dimensions, not as one number.

Our record of November is a small case of exactly that. Every measure of the model was good: eight networks, each standing 20 times of 20. The system lasted six seconds. The cause was not in the network. It was in the first two seconds of the dance the network had been given, a pose with locked knees. It sat between two layers, where a test of one layer does not look.

This is Paqarina on the paper's nine dimensions. The fit is loose in places: they were written for robots that fetch and tidy, not for robots that dance.

DimensionWhat we have measuredWhat is missing
KnowledgeHow close a novel suri, written by the model, is to real suri: 8.1 cm per pointDances it has never seen
UnderstandingThe robot reads 154 numbers about its own joints, its tilt and the danceIt does not see the stage, the partner or the person
ReasoningA plan for a score: 15 pieces of 15 in the dance asked, and 1.69 m reached where 1.5 was askedChanging the plan after a fall
Tool useThe tools of the chain are run by peopleThe robot calls none
Memory137 takes of our own, 4.4 hoursThe robot remembers nothing from one run to the next
Cross-embodimentThe same 26 points on an avatar and on the robot, fitted within 2.7 cm. The ankle at its limit 24 per cent of the timeA second robot
SafetyFalls counted in simulation. On the floor, a hoist and a personA robot that notices the risk itself
Efficiency and costMore than 250 dollars in all. More than nine hours of practice for each dance. Four minutes for the plannerThe cost of a decision made during the dance
User intentThe choreographer's score, in wordsThe dancers' judgement of what comes out

What dance adds to the proposal

Three things matter for dance and are thin in the paper.

The first is time. The paper notes that System 2 takes seconds to decide while control runs in thousandths of a second. A dance step lasts less than a decision. For dance the deciding has to happen between two phrases, or before the music starts. That is why our plan is made ahead.

The second is the body in motion. The paper names dynamic control and walking as open problems. Dance lives there: one foot in the air, the hips turning, the weight never still.

The third is whose dance it is. The nearest dimension is user intent, and it assumes one user. A dance of a community has several: the choreographer, the dancers, and the community the dance belongs to.

What we ask the robot to do

By layer, this is what we ask:

LayerWhat we askSize
System 0Hold 29 angles, 50 times a second, and stay on its feetFrom 31,088 to 46,313 angles for one dance
System 1One dance at a time8 dances, from 21 to 32 seconds
System 2Nothing. A person decides0 decisions made by the robot

Thanks

Paqarina is a Mozilla Fellowship project.

Thanks to Dr Pan Hui, Professor Xin Tong and Jie Yu for allowing us to test our policies on the robot.

What the record shows

MeasureValue
The paperEnvisioning Agentic Robotics: A New Era of Robotic Systems, 13 November 2026, 12 authors
Layers of the proposal that run on our robot2 of 3
Skills on the robot8 dances
Decisions the robot makes0
Parts of a System 2 that Paqarina has, off the robotMemory, planner, verifier
Level reached in simulation1: one whole dance, 8 of 8
On the real robot6 seconds of one dance
Evaluation dimensions in the paper9

What remains open

Whether the verifier can move from the laptop to the moment before the dance: is the robot standing, and is the first pose one it can take. Whether a decision can be made between two phrases without missing the beat. Whether the dancers' judgement can be written as one of the dimensions. The layer that decides is still a person with a remote control. The first thing that person could hand over is a check: whether the robot is standing before the dance goes in.