Humanoid Behavior Foundation Models (BFMs) can execute impressive motions, but downstream applications such as loco-manipulation demand high whole-body positional accuracy. Existing BFMs suffer from a systematic command–execution gap, where executed motions deviate from desired ones. We show that additional test-time compute can improve execution without retraining the BFM. First, we introduce Test-Time Command Optimization (TTCO), which adapts the input command to the specific BFM to produce an executed trajectory closer to the desired motion. TTCO finds such commands by sampling-based optimization through closed-loop simulations of the BFM and robot dynamics. Across three BFMs on the AMASS dataset, TTCO reduces mean global body-position error by 79–84% in simulation. Second, we analyze test-time compute scaling and derive an approximate compute-optimal strategy that can sustain performance gains as the compute budget increases. Third, TTCO can generate BFM-specific action labels from recorded robot trajectories without requiring the original teleoperation body-motion commands. Simulation and real-robot experiments show that these labels are effective for training high-level loco-manipulation policies. This method also supports reusing existing demonstrations across BFMs.
Current humanoid BFMs exhibit systematic deviations between desired and executed motions, a discrepancy we call the command–execution gap. A robot can complete a movement while its body drifts from the intended global position or its hands and feet miss the intended contact locations.
At inference time, the desired motion need not be supplied directly as the command to the BFM. TTCO finds adapted commands through closed-loop simulations of the BFM and robot dynamics. Across three pretrained BFMs and 12,273 AMASS reference motions, TTCO reduces mean global body-position error by 79–84% in simulation.
| BFM | Nominal | Optimized |
|---|---|---|
| HEFT | 0.271 | 0.057 |
| SONIC | 0.615 | 0.127 |
| TWIST2 | 1.095 | 0.173 |
The gains from command optimization depend on both the amount of compute and its allocation across candidates, optimization iterations, rollout horizon, and replanning. A surrogate cost model fitted to limited measurements yields a strategy that sustains performance gains as compute increases and generalizes beyond the motions and BFM used for fitting.
Because of the command–execution gap, recorded robot trajectories may not directly serve as action labels for high-level policy learning. We apply TTCO offline to optimize body-motion commands against these trajectories while keeping the target BFM fixed. The optimized commands provide supervision for policy learning and can be re-optimized for a new BFM to reuse the same demonstrations.
| Task | SONIC → SONIC | SONIC → HEFT | ||||
|---|---|---|---|---|---|---|
| Original | Ours | State | Original | Ours | State | |
| Cup in basket | 55.0 | 55.0 | 0.0 | 0.0 | 40.0 | 13.3 |
| Parcel bag on shelf | 83.3 | 91.7 | 10.0 | 6.7 | 55.0 | 20.0 |
| Basket on shelf | 53.3 | 50.0 | 0.0 | 0.0 | 50.0 | 8.3 |
| Bottle in drawer | 50.0 | 60.0 | 0.0 | 0.0 | 50.0 | 6.7 |
| Mean | 60.4 | 64.2 | 2.5 | 1.7 | 48.8 | 12.1 |