Berkeley’s TANGO lets a humanoid robot navigate with its whole body, not just a map

A University of California, Berkeley and Princeton team has developed TANGO, an AI navigation system that lets a humanoid robot plan movement for all 29 of its joints. Tested on a Unitree G1, it reduced collisions in cluttered environments and completed narrow passages, low obstacles and floor obstacles without real-world navigation training data.

Berkeley BAIR Watch analysis

What happened

The research, reported by Tech Xplore and described in a paper posted to arXiv, treats a humanoid as more than a wheeled robot with legs attached. TANGO takes a natural-language instruction and RGB camera images, then produces whole-body motion rather than a simple two-dimensional route.

Its system combines a vision-language model, a flow-matching action expert that generates short motion sequences, and a motion tracker running at around 200Hz. The researchers trained it with synthetic data: roughly 65,000 verified robot trajectories generated through planning, editing and physics simulation, using about 211 GPU-hours.

On a cluttered navigation benchmark, the team’s initial real-robot tests put TANGO’s collision rate at about 10%, compared with roughly 16% for its strongest modular baseline. When the same model was restricted to predicting a flat 2D path, its real-world success rate fell from about 53% to about 27%.

The G1 also completed 30-metre routes through an office, including side-stepping through narrow spaces, bending under obstacles and stepping over objects. The researchers say this transfer happened zero-shot, without real-world navigation training data.

Key findings

• Whole-body planning
  The robot plans how its arms, torso and legs should move, not merely where its feet should travel.
• Synthetic training data
  Simulation produced tens of thousands of collision-checked trajectories without laboriously recording every manoeuvre by hand.
• RGB-only tests
  TANGO used ordinary camera images while the comparison system also had LiDAR, yet the reported collision rate was lower.
• Real-world transfer
  The G1 navigated office routes and awkward obstacles without additional real-world navigation training.

Why it matters

Humanoid navigation is a geometry problem with elbows. A route that looks clear on a map may be impossible for a tall, articulated machine whose body changes shape as it walks. TANGO’s central idea is that the robot’s posture is part of navigation itself, not a downstream detail for a controller to sort out.

That matters for homes, offices and care settings, where robots will face narrow gaps, clutter, poor lighting and objects that move. Synthetic data could also make training more scalable, although success in an office demonstration is not the same as reliable operation around people.

Our read

This is a meaningful robotics result because it connects language, visual understanding and physical movement in one system, with a measurable comparison against a modular baseline. The reported collision improvement is promising, and the 2D-path ablation makes the argument unusually clear: for humanoids, “go there” is not enough. The machine needs to know how it will fit while getting there.

The important test now is reproducibility beyond the team’s benchmark and office route. The researchers acknowledge that the low-level tracker struggles with harder terrain such as stairs, while RGB-only perception can suffer in ambiguous or poorly lit scenes. The dataset, model and synthesis pipeline are expected to be released openly, which gives other researchers a chance to check whether the clever bit travels.

What to watch

• Open release:
  whether the promised TANGO code, model and dataset arrive in usable form.
• Harder terrain:
  performance on stairs, uneven ground and changing surfaces.
• Human environments:
  safety and reliability around moving people rather than static office clutter.
• Loco-manipulation:
  whether the system can progress from avoiding obstacles to opening doors or moving objects.

Discussion spark: Should humanoid robots be judged first by how intelligently they plan movement, or by whether they can prove reliable safety around unpredictable people?

Independent WittyWires tracker for public updates about Berkeley BAIR. Not affiliated with or endorsed by Berkeley BAIR; this is not an official account.