Berkeley's TANGO Ships 29-DoF Humanoid VLA on Unitree G1

The system predicts 29 joint angles at every step, straight from a natural-language instruction and two egocentric RGB feeds. That is the pitch of TANGO, a whole-body vision-language-action model for humanoid navigation from researchers at UC Berkeley, Peking University, Tsinghua, HKU and Princeton. The stated goal is language-guided traversal of cluttered indoor scenes, with the humanoid sidling, ducking or striding around obstacles rather than treating the problem as a 2D path plan.

The architecture stacks a 7-billion-parameter Qwen2.5VL backbone, warm-started from InternVLA-N1, on top of a flow-matching diffusion action expert. Underneath sits a lower-level tracker the authors call SONIC that runs at roughly 200Hz on-robot. Training was done entirely in simulation on 64,633 synthesized trajectories drawn from 578 scenes, at a cost of 896 A100 GPU-hours plus 211 RTX PRO 6000 hours for data generation.

The paper reports 51.30% success on VLNVerse-seen and 43.75% on unseen, with a 9.90% collision rate on the cluttered variant, better than the strongest modular baseline the authors tested. Deployed zero-shot on a Unitree G1, TANGO landed 13 of 15 trials on short-horizon 2D navigation, 14 of 15 on a roughly 30-metre long-horizon run, and 13 of 15 in cluttered 3D scenes with an obstacle, averaging 0.20 collisions per trial.

The authors write: "we deploy TANGO zero-shot on a Unitree G1 humanoid robot, and observe robust language-guided traversal in cluttered real-world scenes without training on any real-world navigation data." An ablation that removed the real-time chunking component cost 32.81 percentage points of success rate, which the paper treats as evidence that motion continuity across chunks is the load-bearing piece.

It lands amid a busy stretch for robotics coverage on our tracker, including this week's CSRC bar raise for humanoid IPOs that pointedly names Unitree.