GRAFT: Geometry-Aware Residual Adaptation for Foundation Trackers
A preprint from Xiaomi Robotics Lab reports that a small add-on network raised SONIC, a frozen humanoid motion tracker, from 3 to 39 successes in 40 hard-terrain simulation runs. On a real Unitree G1 with onboard depth, the adapted tracker crossed a 0.30 m platform and steps of 0.10 m and 0.17 m that its unmodified version could not, using the same motion references.
Can a humanoid walk stairs, gaps and platforms by adding a small terrain correction to a frozen whole-body tracker, without retraining the tracker itself?
Behavior foundation models are single policies that follow any kinematic reference, whether a recorded clip, a teleoperator or a VLA model. They are trained mostly on flat-ground motion, so they have no way to change foot placement or swing height for the ground ahead.
Many terrain-aware approaches retrain or fine-tune the tracker together with a perception stack. That costs data and compute, and it ties ordinary tracking to the perception pathway, so the tracker can no longer be swapped or updated as a separate part.
GRAFT keeps the tracker's parameters and reference interface untouched. A light adapter adds terrain-conditioned corrections only to lower-body execution targets. It is trained with task and foot-geometry rewards, penalised for corrections on flat ground, and then distilled from privileged terrain data to onboard depth.
Across ten sessions per terrain cell, successes without and with GRAFT were 3 to 39 of 40 on SONIC, 3 to 34 on Humanoid-GPT, 8 to 39 on HoloMotion and 2 to 20 on ScaleBFM. On a composite course, the mean correction was 1.529 on obstacles against 0.408 on flat segments, peaking at 3.488 on the gap. Waist tracking error rose from 1.18 degrees on flat ground to 3.17 degrees on obstacles. On a Unitree G1, the adapted SONIC crossed a 0.30 m platform and a 0.10 m step where the frozen base trod in place, and also handled a 0.17 m step and a push-cart scene.
This is a preprint. The success rates are simulation figures from small samples, n = 10 per cell. The hardware results are demonstrations with no reported trial counts. Gains are uneven: ScaleBFM fell from 1 of 10 to 0 of 10 on down stairs, and Humanoid-GPT reached only 5 of 10 on the platform. All trackers used SONIC dataset references, and the authors say the curves are not a ranking of trackers. Each tracker needs its own adapter.
For humanoid makers that license or share a motion tracker, this suggests terrain handling could ship as a separate module rather than a full retrain. It is notable that the lead lab belongs to Xiaomi, a phone and EV maker with a robotics unit, and that the test robot is a Unitree G1 running a tracker published elsewhere. Near term the value is in research and pilot deployments, since hardware reliability has not been measured.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.