HALO humanoid trained only in sim hits 91% in real homes
Siyuan Hu's team says the policy went straight from 1,000 procedurally generated homes to six real environments with no fine-tuning.
Siyuan Hu has released HALO, a perceptive humanoid motion system trained entirely on synthetic data, with the paper, code and project page all live.
Trained entirely on synthetic data from 1,000 procedurally generated homes. Zero-shot deployment in 6 unseen real environments. 91% success rate across 360 physical trials. Fully onboard.
The training set is 1,000 procedurally generated homes. No real-world data, no real-world fine-tuning. The team then deployed it zero-shot (straight from simulation, with no additional training on the real robot) in 6 unseen real environments and reports a 91% success rate across 360 physical trials. It runs fully onboard, so no off-robot compute and no remote brain doing the thinking.
Why the setup matters more than the number
Sim-to-real for whole-body humanoid motion is the part of the stack where teams usually admit they had to collect real data in the end. Generating a thousand homes in a simulator is cheap compared with teleoperating a humanoid through a thousand real ones, so if a policy trained that way holds up in houses it has never seen, the data bottleneck looks smaller than it did. Six environments and 360 trials is a real number of attempts, not a single hero clip.
It is also the team's own evaluation on its own trials. Nobody outside has run HALO yet, and the sources do not say what the failures looked like or how hard the six environments were. That is the usual gap in robot results right now, and it is exactly the gap projects like RoboColosseum are trying to close by running models head to head on shared hardware.
The wider mood check came from Chris Paxton the same week, who posted that there is still no true general purpose robotics model, not even from OpenAI. HALO is not claiming to be one. It is claiming that a specific capability, perceptive humanoid motion through cluttered homes, can be learned without touching real data first. Treat the 91% as a strong first-party result and wait for someone else to try it.
There remains no true general purpose robotics model -- not even from OpenAI
Sound like you were at the demo
One email a day. Under three minutes. The demos, the deployments and the part the press release left out.