Humanoid robot control has a shape problem. A wheeled robot is a disc that has to find a gap wider than itself; a humanoid is a tall, articulated body whose silhouette changes with every step, and which can shrink itself by tucking an arm or turning sideways. Almost every navigation system built so far throws that flexibility away, planning a flat two-dimensional path and asking the legs to follow it.
A team from the University of California, Berkeley and Princeton University has published a model that does not. TANGO takes a natural-language instruction and a camera feed and outputs commands for all 29 joints of a humanoid at once. Our AI models and tools hub tracks the vision-language model families this is built on, and our ML model development page covers how systems like it get trained and shipped.
This article sets out the architecture, the way the training data was manufactured, the measured results in simulation and on a real robot, and the limits the authors state plainly.
Table of contents
- What the TANGO Humanoid Robot Control Model Actually Does
- How the Humanoid Robot Control Training Data Was Manufactured
- The Numbers Behind the Humanoid Robot Control Result
- Why Joint Space Is the Whole Humanoid Robot Control Argument
- Where This Humanoid Robot Control Approach Still Breaks
- Why Cluttered Space Is the Real Humanoid Robot Control Test
- How TANGO Compares With Other Humanoid Robot Control Work
- What the Plan-Edit-Track Idea Means Beyond Humanoid Robot Control
- What Comes After Navigation for Humanoid Robot Control
- What Humanoid Robot Control Teams Should Take From TANGO
- How to Read a Humanoid Robot Control Demo Video
- Frequently Asked Questions About Humanoid Robot Control and TANGO
- References
What the TANGO Humanoid Robot Control Model Actually Does
The humanoid robot control paper is “TANGO: Humanoid Navigation in Cluttered Environments with a Whole-Body Vision-Language-Action Model,” posted to arXiv under DOI 10.48550/arxiv.2609.09158.
Language in, joint commands out
TANGO accepts a natural-language instruction and RGB camera input, and emits whole-body motion for the robot’s 29 joints. It is not producing a steering command for something else to execute; the joint targets are the output.
A vision-language model does the interpreting
The front of the stack is a vision-language model that reads the scene alongside the instruction. That is what lets the system connect “go to the desk behind the boxes” to the pixels in front of it.
A flow-matching action expert produces the motion
Behind it sits a flow-matching action expert that generates motion in real-time chunks rather than one pose at a time. Chunked generation is what makes the output rate practical for humanoid robot control rather than merely feasible offline.
An off-the-shelf tracker executes at 200 Hz
The generated motion is handed to a standard motion tracker running at roughly 200 Hz — a command every 5 milliseconds. Notably, the tracker is not novel; the contribution sits above it.
The team behind it
Anqi Li led the work, with Yuxin Chen, Zhaobo Li, Zhuo Cao, Junli Ren and Masayoshi Tomizuka at UC Berkeley, and Dhruv Shah of Princeton as co-senior author.
| Layer | Role |
|---|---|
| Vision-language model | Reads the scene and the instruction together |
| Flow-matching action expert | Generates real-time motion chunks |
| Motion tracker | Off-the-shelf, executes at roughly 200 Hz |
| Output space | All 29 joints, not a 2D path |
| Sensing at test time | RGB cameras only |
| Hardware tested | Unitree Robotics G1 |
How the Humanoid Robot Control Training Data Was Manufactured
Whole-body humanoid robot control demonstrations through cluttered space are expensive to collect by teleoperation. The team’s answer was to synthesise them, and the pipeline is arguably the more transferable contribution.
Plan a safe route first
The Plan-Edit-Track pipeline starts by planning a collision-free route through a cluttered simulated scene, which gives a clean skeleton for the motion to hang on.
Edit the walk to add whole-body behaviour
The plain walking motion is then edited to insert the behaviours that make humanoid robot control worth the complexity: tucking arms in, crouching under overhead obstacles, stepping over objects on the floor, and turning sideways to fit a narrow gap.
Replay it in physics and throw away the failures
Each edited trajectory is replayed in a physics simulator. Anything that collides or falls is discarded, so what survives is verified rather than merely generated.
The resulting dataset
That process produced roughly 65,000 verified trajectories using about 211 GPU-hours of compute — around 308 usable trajectories per GPU-hour, or roughly 11.7 seconds of GPU time per surviving trajectory.
Why this matters more than the model
A humanoid robot control dataset of that size collected by hand would take months of operator time and a tolerant hardware budget. The team has said the data-synthesis pipeline, the dataset and the model will all be open-sourced.
The Numbers Behind the Humanoid Robot Control Result
Two sets of humanoid robot control figures carry the paper: one from simulation and one from a real machine.
Roughly double in simulation
On the simulated benchmark, joint-space control succeeded 53% of the time against 27% for the restricted two-dimensional path approach. That is 1.96 times as often, or 26 percentage points better.
A 37.5% relative cut in collisions
On the real robot, collision rate fell from about 16% for the baseline to about 10% with TANGO. Six percentage points off sixteen is a 37.5% relative reduction, which is the more honest way to state it.
Thirty metres, zero real-world training
The robot ran 30-metre routes through a cluttered office, side-stepping narrow passages, bending under obstacles and stepping over objects on the floor, with no real-world training data at all. That is a zero-shot transfer from simulation.
Cameras only, against a baseline with LiDAR
TANGO used RGB cameras alone. The baseline it beat had LiDAR available as well, which makes the comparison harder on TANGO rather than easier.
Twenty-nine dimensions against three
A two-dimensional path is essentially three numbers: position and heading. Humanoid robot control in joint space is 29. The claim is that the extra dimensions are where the clearance comes from.
| Measure | Value | Where |
|---|---|---|
| Joint-space success rate | 53% | Simulation |
| 2D-path success rate | 27% | Simulation |
| Relative improvement | 1.96x | 53 / 27 |
| Baseline collision rate | ≈16% | Unitree G1 |
| TANGO collision rate | ≈10% | Unitree G1 |
| Relative collision cut | 37.5% | 6 / 16 |
| Training trajectories | ≈65,000 | Synthesised |
| Compute to produce them | ≈211 GPU-hours | Simulation |
Why Joint Space Is the Whole Humanoid Robot Control Argument
Shah’s framing of the humanoid robot control problem is the clearest statement of why this is not an incremental result.
The body is not a disc
“A humanoid is a tall, wide, articulated body whose shape changes continuously as it moves,” Shah said. Every assumption in classical mobile robot navigation is built on a footprint that does not change.
The thesis, stated bluntly
“Reasoning in joint space is not a detail; it is the point,” Shah said. A system that plans in two dimensions and then asks a controller to follow the line has already discarded the behaviours that let a body fit through a gap.
Clearance is a posture, not a measurement
Whether a humanoid fits through a doorway depends on what it does with its arms and torso. Treating clearance as a fixed number makes half the reachable space look impassable.
Which is why the baseline fails more often
The restricted two-dimensional path approach succeeding only 27% of the time in cluttered scenes is not a tuning failure. It is the cost of an abstraction that cannot express crouching.
And why 29 joints is not overkill
Humanoid robot control that outputs all 29 joints is harder to learn and harder to stabilise. The 26-point gap is the evidence that the difficulty buys something.
Where This Humanoid Robot Control Approach Still Breaks
The authors are direct about the humanoid robot control limits, and the limits are specific enough to be actionable.
The tracker becomes the bottleneck
On complex terrain such as stairs, the low-level motion tracker — the off-the-shelf component — is the constraint. The model can ask for motion the tracker cannot reliably deliver.
RGB-only has predictable failure conditions
Relying on cameras alone means poorly lit rooms and visually ambiguous scenes are hard. The LiDAR the baseline carried exists for exactly those cases.
A 10% collision rate is still one in ten
Cutting collisions by 37.5% is a real result, and a humanoid robot control system that clips something on one run in ten is not yet deployable near people without supervision.
Success at 53% is a benchmark number, not a product number
Just over half in simulation is a strong comparative result and a weak absolute one. The right reading is “better than the alternative,” not “solved.”
One robot, one embodiment
Every real-world figure comes from a Unitree G1. Transfer to a differently proportioned humanoid with different joint limits is an open question.
Why Cluttered Space Is the Real Humanoid Robot Control Test
Open-floor humanoid robot control demos have been convincing for years. Clutter is where the useful work is, and where almost every deployment plan currently stalls.
Human buildings are built to human clearances
Corridors, doorways and the gaps between desks were sized for people who duck, turn and tuck without thinking. A machine that needs a wider margin than a person cannot use most of the building.
Warehouses are not the hard case
Purpose-built logistics floors are already served by wheeled and rail-guided machines. The argument for legs only starts where the environment cannot be redesigned around the robot.
Retrofitting a site is the alternative cost
If a humanoid robot control system cannot fit a gap, the other option is widening the gap. That cost lands on the site owner and is usually the reason a pilot does not become a rollout.
Overhead obstacles are the overlooked half
Pipework, shelving and signage sit at head height in exactly the buildings where robots are proposed. Crouching under them is a behaviour a two-dimensional planner has no way to express.
Contact is the safety question
Near people, the metric that matters is whether the machine touches anything it did not intend to. That is why collision rate rather than traversal speed is the number this paper leads with.
How TANGO Compares With Other Humanoid Robot Control Work
Narrow-terrain humanoid robot control is an active field, and TANGO’s position in it is specific.
Proprioception-only locomotion took a different route
Separate work has shown humanoids crossing narrow beams and extreme terrain using body-sense alone, trained with reinforcement learning and no cameras at all. That solves balance; it does not solve knowing what is ahead.
Planning-based methods solve geometry precisely
Whole-body planners using differentiable collision detection can find routes through tight configuration spaces with guarantees. They need a model of the scene, which is the thing a camera-only system does not have.
TANGO’s distinctive claim is the combination
Language instruction, raw camera input and whole-body output in one learned policy is the unusual part. Each piece exists elsewhere; the end-to-end path is what the paper is arguing for.
The comparison is run on fair ground
TANGO beating a LiDAR-equipped baseline while using cameras only is the sort of comparison that resists the usual objection, because the handicap runs the wrong way for the authors.
Open-sourcing changes what happens next
A released dataset and pipeline mean the next humanoid robot control paper can attack the same benchmark rather than inventing its own, which is how a field starts making comparable claims.
What the Plan-Edit-Track Idea Means Beyond Humanoid Robot Control
The data pipeline is the part of this paper most likely to be reused by people who never touch a humanoid.
Editing a known-good trajectory beats generating from scratch
Starting from a planned, feasible route and perturbing it keeps the synthetic data near the manifold of things that actually work. Pure generation wanders off it.
Physics is a cheap, honest label
Running each candidate through a simulator and discarding falls gives a verified label with no human in the loop. The filter is the supervision.
The cost structure is unusually transparent
Two hundred and eleven GPU-hours for 65,000 usable trajectories is a number any team can compare against its own teleoperation budget, which is rare in humanoid robot control papers.
It scales with scene variety, not operator hours
Adding harder environments costs compute rather than staff time, which changes what a small group can attempt.
The same shape works for manipulation
Plan a grasp, edit the approach, replay in physics, keep the successes. Nothing about the pattern is specific to legs, which is presumably why the team names loco-manipulation as the next target.
What Comes After Navigation for Humanoid Robot Control
The team has said where the humanoid robot control work goes next, and the direction is more demanding than what has been achieved.
From avoiding obstacles to moving them
The stated next step is interaction: opening doors and moving obstacles rather than routing around them. That turns a navigation problem into a manipulation problem mid-stride.
Loco-manipulation is the stated goal
The authors describe working toward a general loco-manipulation model — walking and handling objects under one policy rather than two stacked systems.
The tracker will have to improve first
If the low-level tracker already limits humanoid robot control on stairs, it will limit it far more when the robot is also carrying or pushing something.
Sensing will probably have to widen
Camera-only sensing is an elegant constraint for a navigation paper. Contact-rich interaction usually needs force feedback, which is a different sensor budget.
The open-source release is the enabler
If the pipeline, dataset and model land publicly as promised, other groups can extend the edit step with manipulation behaviours without rebuilding the whole synthesis stack.
| Claim | Evidence status |
|---|---|
| Whole-body beats 2D-path planning | Measured, 53% against 27% in simulation |
| Transfers zero-shot to hardware | Demonstrated on 30-metre office routes |
| Cameras suffice without LiDAR | Shown here; fails in poor light by the authors’ own account |
| Works on stairs | Explicitly limited by the tracker |
| Generalises across humanoids | Untested; one robot model |
| Can manipulate objects | Stated as future work |
What Humanoid Robot Control Teams Should Take From TANGO
For anyone building or buying in this space, there are four or five transferable lessons that do not depend on adopting the model.
Synthesised data with a physics filter is credible
Generating trajectories and discarding the ones that fall is a cheap way to get verified demonstrations. The 211 GPU-hours figure is the number to benchmark your own data budget against.
Report the relative figure, not just the absolute
A drop from 16% to 10% is a 37.5% relative cut. Both numbers belong in any honest claim, and a vendor quoting only one of them is choosing which impression to leave.
Handicap yourself in comparisons
Beating a LiDAR baseline with cameras only is more persuasive than beating a weaker system with better sensors. Build evaluations that can embarrass you.
Name the bottleneck component
Saying the off-the-shelf tracker is the constraint on stairs is a precise, testable admission. It tells the next team exactly where to work.
Judge humanoid robot control by clearance, not speed
The interesting metric in cluttered space is whether the machine gets through without contact, not how fast it crosses open floor.
How to Read a Humanoid Robot Control Demo Video
Humanoid robot control marketing is dominated by clips, and a clip is engineered to hide exactly the things this paper measures.
Ask how many takes it took
A success rate is a statement about the runs you did not see. TANGO’s published 53% and 10% figures exist precisely because someone counted the failures.
Ask what the robot could see
A humanoid robot control demo running on a pre-scanned map of the room is solving a different problem from one running on live camera input. Both look identical on video.
Ask whether the route was known
Zero-shot transfer into an unseen cluttered office is a much stronger claim than a rehearsed path through a fixed obstacle course, and the difference is invisible on screen.
Ask what happens at the edges
Poor light, reflective floors, glass partitions and stairs are the conditions that break camera-only systems. A demo shot in an evenly lit lab has quietly excluded all of them.
Ask who the baseline was
Beating an unnamed prior system means little. TANGO’s comparison is against a LiDAR-equipped baseline on the same hardware, which is the kind of detail that makes a number worth repeating.
Frequently Asked Questions About Humanoid Robot Control and TANGO
What is TANGO?
A whole-body vision-language-action model for humanoid navigation in cluttered environments, from UC Berkeley and Princeton, published on arXiv under DOI 10.48550/arxiv.2609.09158.
How is it different from normal robot navigation?
It outputs commands for all 29 joints from language and camera input, rather than planning a two-dimensional path for a controller to follow.
How much better did it perform?
53% success against 27% in simulation — 1.96 times as often — and a real-robot collision rate of about 10% against a baseline’s 16%, a 37.5% relative reduction.
What robot was it tested on?
A Unitree Robotics G1, running 30-metre routes through a cluttered office with no real-world training data.
How was the training data produced?
Through a Plan-Edit-Track pipeline that plans a safe route, edits the walk to add whole-body behaviours, replays it in physics and discards collisions and falls — roughly 65,000 verified trajectories from about 211 GPU-hours.
What are the main limitations?
The off-the-shelf motion tracker limits performance on complex terrain such as stairs, and RGB-only sensing struggles in poorly lit or visually ambiguous scenes.
Will the work be released?
The team has said the data-synthesis pipeline, the dataset and the TANGO model will be open-sourced.
References
TANGO: Humanoid Navigation in Cluttered Environments with a Whole-Body Vision-Language-Action Model
Humanoid robots navigate narrow gaps and obstacles with whole-body AI control
Humanoid Whole-Body Locomotion on Narrow Terrain via Dynamic Balance and Reinforcement Learning
Berkeley Artificial Intelligence Research
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
