A single humanoid policy reported higher full-episode success than LHM-Humanoid when moving more than one object through cluttered indoor scenes, including a set of tasks held out from training. In the 350-task comparison, full-episode success was 80.4% for the proposed policy and 71.1% for LHM-Humanoid. On 66 unseen tasks, the corresponding figures were 78.0% and 61.6%.
The paper is an arXiv preprint dated 6 October 2026. It examines long-horizon whole-body loco-manipulation, meaning the controller must coordinate walking and object handling across consecutive transport stages. The two-object benchmark contains 350 training scenes and 66 held-out scenes.
Three changes to the training setup
The method uses one shared policy trained across all transport stages at once, an approach the paper calls Parallel Training. Dynamic Starting gradually replaces fixed proxy starting points with handoff states taken from upstream rollouts. Reward Gating zeros the remaining reward for later stages if the preceding object has been displaced beyond a threshold, with the stated aim of protecting placements that have already been completed.
The benchmark counts a transport cycle as successful when the object root ends within 0.2 metres of its target. The reported measures cover the first and second cycle, the complete episode, and final object-to-goal distances, using the labels Succ1, Succ2, SuccAll, Dist1 and Dist2.
Held-out scenes put the gap in sharper focus
On the 350-task comparison, the full system reported 96.6% for the first cycle, 81.3% for the second and 80.4% for the whole episode. LHM-Humanoid reported 87.4%, 72.6% and 71.1%. The final distances were also lower for the full system: 0.06 and 0.38 metres, compared with 0.28 and 0.50 metres.
On the 66 unseen-task benchmark, the full model reported 97.1%, 79.3% and 78.0% for Succ1, Succ2 and SuccAll. LHM-Humanoid's SuccAll was 61.6%, a reported gap of about 16 percentage points. The ablation without Dynamic Starting scored 63.5% on SuccAll, while the version without Reward Gating reported 89.1%, 72.3% and 66.4% for the three measures.
What the component checks show
On seen tasks, removing either component was accompanied by lower full-episode scores. Without Dynamic Starting, the reported SuccAll was 67.7%, despite a 92.7% first-cycle score. Without Reward Gating, SuccAll was 69.6%. These are computational ablations, so they show how the reported configurations differed, not that either component alone caused the gap.
Longer chains still expose the limit
The advantage did not remove the difficulty of extending the chain. In a five-object test, success was 82.2% at the third stage, 66.5% at the fifth and 60.2% when the whole sequence had to succeed. The paper describes this decline as graceful and says the full model stayed ahead of the compared methods at each stage. Without Dynamic Starting and without Reward Gating, full-episode success was 49.8% and 38.9%, respectively.
The failure analysis points to a larger gap between the first and second cycles when later navigation begins from an arbitrary post-task configuration. It also reports frozen-action failures during unstable grasps, and links uniformly allocated scenes to difficult clutter cases that may be under-optimized.
Training curves were relatively consistent across the reported seeds: final SuccAll averaged 87.6% plus or minus 0.3 percentage points, while Succ1 averaged 92.5% plus or minus 0.3 percentage points. Those spreads are standard deviations, not confidence intervals, so they describe variation across the reported runs rather than uncertainty around a population estimate.
A perception test, not a deployment result
The paper also reports two perception-conditioned students. The RGB student recorded 93.5% for Succ1, 76.8% for Succ2 and 75.6% for SuccAll. The depth student recorded 93.7%, 77.6% and 73.7%, respectively. Dist1 was 0.07 metres for both students; Dist2 was 0.50 metres for RGB and 0.52 metres for depth.
Those figures are not physical-robot validation. The evidence is limited to simulation-based benchmark experiments, algorithmic comparisons, five-object extension tests and exploratory VLA distillation. The work therefore does not establish performance under real-world sensing, contact variation or safety constraints, and the five-object results do not show how reliably the approach would scale to longer horizons.
The evidence remains bounded
The results are best read as evidence from a simulated benchmark, not as a demonstration of deployment readiness. Data and code were not open with the submission; the checklist says the dataset and code will be released upon acceptance.
Paper data and sources
Original title: Humanoid Horizon: Extending Task Horizon in Whole-Body Loco-Manipulation via Parallel Training, Dynamic Starting, and Reward Gating
Authors: Haozhuo Zhang, Qiang Zhang, Jian Tang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-10-06
DOI: Not available
Original paper · Full text