acceptodds
Under review as a conference paper at ICLR 2027

Test-Time Compute for Humanoid Behavior Foundation Models

Abstract

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. The project page is available at https://bfm-ttc.github.io/

Primary area:
applications to robotics, autonomy, planning
Keywords:
Humanoid Robot, Behaviour Foundation Model, Test-Time Scaling, Sampling-based Optimization

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_zS5DOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Test-Time Command Optimization (TTCO) to improve motion-tracking accuracy of pretrained humanoid Behavior Foundation Models (BFMs) without retraining. By optimizing motion commands through simulation-based rollouts using sample-based methods, TTCO reduces tracking errors. The paper further studies compute-optimal allocation and applies TTCO to generate action labels for downstream loco-manipulation and cross-BFM demonstration reuse.

Critical strengths
  • The paper tries to address a practical and important problem, formalizes it, and proposes a test-time optimization method rather than training a stronger BFM.
  • TTCO reduces mean global body-position errors by 79–84% across three BFMs on 12,273 AMASS motions, demonstrating strong effectiveness and generality.
Critical weaknesses
  • The method introduces additional computational cost at test time, which raises concerns about real-time inference.
Questions for authors
  • Is the computational cost of test-time optimization acceptable in practice? Could the authors report per-planning-step latency and analyze the trade-off between wall-clock runtime and tracking performance under different compute budgets?
  • Can the optimized data be used to directly reduce the BFM's tracking errors? For example, could the authors collect (command, action) pairs and finetune the BFM using behavior cloning (BC)?
  • Does the method require additional steps when the robot interacts with objects?
  • Table 1 shows increased joint acceleration after optimization, and the authors acknowledge potential motion jitter. Could the authors provide a quantitative analysis of command discontinuities and execution stability, particularly at replanning boundaries? How can such instability be mitigated?
Reviewer_NPTqOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper observes that when the desired motion is directly used as the command to a humanoid Behavior Foundation Model (BFM), the executed motion systematically deviates from the desired one. To address this, the authors keep the BFM unchanged and design an outer sampling-based optimization loop in simulation (receding-horizon MPC with CEM) that searches for better commands, so that the BFM tracks the desired motion more accurately (reducing tracking error by roughly 80% in simulation). The authors further analyze how test-time compute scales and should be allocated in this optimization process, and use the optimized commands as training labels for high-level policies, enabling demonstration data to be transferred across different BFMs.

Critical strengths

S1. General idea with consistent large-scale gains. A major problem with current BFMs is their limited tracking accuracy, especially for manipulation tasks. The core idea of this work addresses this problem well. TTCO requires no retraining of the BFM and can be applied to different BFMs and motions. Across 12,273 motions and three different BFMs, it reduces body-position error by 79–84% in simulation (Table 1). This consistency suggests that the command–execution gap is a general issue of current BFMs and that the method is broadly applicable.

S2. Practical reuse of demonstrations across BFMs. By running the optimization offline to generate training labels, the paper allows demonstrations collected with one BFM to be reused for another. On the real robot, re-optimizing SONIC-collected demonstrations for HEFT raises the success rate from 1.7% to 48.8% (Table 3). This addresses a practical bottleneck, since recollecting teleoperation data for every new controller is costly.

Critical weaknesses

W1. Lack of analysis of the BFM's deviations; the approach acts more like a patch. The paper only briefly lists possible causes of the command–execution gap (Sec. 3.1) and does not analyze which factors dominate or how the deviations are structured. TTCO compensates for the gap through per-motion black-box search, so the paper shows that adjusting commands helps but not why. For example, it does not explain why gains differ so much across BFMs (cost-win rate 62% for HEFT vs. 99.7% for TWIST2, Table 1). Analyzing the optimized corrections (e.g., which joints are corrected most) would make the contribution more informative.

W2. Per-motion optimization undermines the "foundation" nature of BFMs. BFMs are appealing because they execute arbitrary motions zero-shot, whereas TTCO requires a separate simulation-based optimization for every motion and an accurate model of the robot and environment. Based on the appendix's wall-clock formula, a 10-second motion takes roughly 98 seconds to optimize under the Table 1 configuration. Although Sec. 4.4 avoids this cost at deployment by optimizing labels offline before training, the tracking setting itself still requires per-motion optimization, which limits its applicability to new motions and environments.

Questions for authors
  1. Why is the problem non-differentiable? Sec. 3.2 states that the full-sequence problem is "non-differentiable" and therefore intractable to optimize directly, but this premise is not explained. Could the authors clarify which component makes the problem non-differentiable in their setting (e.g., contact discontinuities, unreliable gradients over long horizons, or lack of simulator support)?

  2. Label transfer pipeline and analysis of Table 2. The cross-BFM label generation procedure is scattered across Sec. 4.4 and Appendices C.5.1–C.5.4, which makes it hard to follow; a concise description or diagram in the main text would help. In addition, the results in Table 2 are not analyzed: (a) beyond the existence of the command–execution gap, is there direct evidence of why State labels fail? (b) In the same-BFM setting, why does Ours fall well below Original in simulation (56.7 vs. 75.8) but slightly exceed it on hardware (64.2 vs. 60.4)?

  3. Magnitude of the sim-to-real gap. TTCO relies heavily on the simulated dynamics used for optimization, but tracking errors are reported only in that same simulator (Table 1), and real-robot tracking is shown only qualitatively (Fig. 4). Task success rates in Sec. 4.4 are an indirect and coarse measure of whether the optimized commands transfer. Could the authors report direct tracking errors when commands optimized in MuJoCo Warp are executed in a different simulator (e.g., Isaac Sim) or on the real robot, for a subset of the motions in Table 1?

  4. Do policies trained on optimized labels output commands the BFM can execute well? The optimized commands are corrections computed for specific states along each demonstration and are less smooth than the original commands (Sec. 5). At deployment, the high-level policy encounters different states and may drift from the training distribution, so the learned corrections may no longer be appropriate. Have the authors measured the command–execution gap of the policy's outputs at deployment? This might also explain why Ours remains below Original in the same-BFM setting (56.7 vs. 75.8, Table 2).

Minor questions

  • 0.0% entries in Table 2. In every cross-BFM setting (Table 2 and Table 25), all three label types achieve 0.0% success on Open Door and Pick & Place. Could the authors discuss why these tasks fail completely after transfer (e.g., contact-rich manipulation not being captured during label optimization)?
  • Choice of TTCO configurations. The main tracking results use (I, N, H, E) = (4, 80, 40, 20) with K = 640 (Table 1), while label generation uses (7, 39, 37, 5) with a nominal K = 2048 (Sec. 4.4). How were these configurations chosen, and was the Table 1 configuration selected using the surrogate model of Sec. 3.3? Its horizon ratio H/E = 2 is below the point where Fig. 5(c) shows the cost leveling off (H/E = 4).
Reviewer_JcJLOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Test-Time Command Optimization (TTCO) to reduce the gap between desired and executed motions in humanoid behavior foundation models (BFMs). Rather than modifying the pretrained policy, TTCO searches for motion commands that lead to more accurate tracking through closed-loop simulation rollouts. The paper further studies how to allocate a fixed test-time compute budget and fits a surrogate cost model to select optimization configurations. Finally, it applies optimized commands as action labels for downstream loco-manipulation policies and cross-BFM demonstration reuse.

The experiments cover three pretrained BFMs, 12,273 AMASS motions, compute-allocation studies, and simulated and real-robot loco-manipulation tasks. The results suggest that command optimization can substantially improve tracking accuracy and can be useful when the original motion-command labels are unavailable.

Critical strengths

A useful problem formulation. The paper identifies a practical limitation of humanoid BFMs: the desired reference motion is not necessarily the best command for obtaining the intended physical execution. Optimizing the command while keeping the BFM fixed is a simple and potentially useful way to address this issue.

Broad evaluation across pretrained controllers. Evaluating TTCO on HEFT, SONIC, and TWIST2, together with a large set of AMASS motions, provides stronger evidence than a demonstration on a single controller or a few selected motions. The reported reductions in global body-position error are substantial.

A useful investigation of compute allocation. The paper examines the interaction between sampling effort, planning horizon, and replanning frequency instead of treating the test-time budget as a single scalar. The surrogate model also offers a practical way to select configurations without exhaustively evaluating every possibility.

Evidence of downstream utility. The simulated and real-robot loco-manipulation experiments go beyond motion-tracking metrics. In particular, the cross-BFM experiments suggest that optimized labels can help reuse demonstrations when the target controller differs from the source controller.

Critical weaknesses

The novelty relative to existing sampling-based control methods needs to be established more clearly. TTCO uses sampling-based optimization over motion commands with policy-in-the-loop physics rollouts. The paper discusses related work such as Sumo and BFM-Zero, but it remains unclear which aspects of the formulation provide the essential advance beyond applying existing trajectory-optimization techniques to a broader set of pretrained BFMs. A more direct comparison or a clearer isolation of the methodological contribution would strengthen the paper.

The compute-optimal strategy is validated over a limited search space. The surrogate model is fitted using a relatively small set of measured configurations, and the allocation studies use only ten motions. Evaluation on additional motions and BFMs is encouraging, but the results do not fully establish how reliably the fitted model identifies near-optimal configurations under substantially different motion distributions or compute budgets. The reported regret is also relative to the best measured configuration, which is not necessarily the true optimum.

The tracking improvements involve a smoothness trade-off, and the practical compute cost remains unclear. Table 1 reports increased joint acceleration for all three BFMs, including a substantial increase for TWIST2. This matters because improved positional tracking does not necessarily imply more natural or stable execution. In addition, the paper acknowledges that the current implementation is not sufficiently fast for real-time command optimization. The transition from simulated optimization to practical deployment therefore remains an important limitation.

The downstream results are not uniformly positive. In the same-BFM simulation setting, policies trained with optimized labels achieve a mean success rate of 56.7%, compared with 75.8% for policies trained on original commands. This supports the claim that optimized labels are useful when original labels are unavailable, but not that they are generally equivalent to the original supervision. The real-robot and cross-BFM results are more encouraging, although the extent of the benefit varies across settings.

Questions for authors
  1. How does TTCO differ conceptually and empirically from Sumo and BFM-Zero? Can the authors isolate the benefit of the proposed formulation from the benefit of simply increasing the number of policy-in-the-loop rollouts?

  2. How sensitive is the compute-optimal strategy to the motion set used to fit the surrogate model? In particular, does the strategy remain competitive when evaluated on substantially different motion categories or on a BFM not represented in the fitting data?

  3. Can the authors quantify the trade-off between tracking accuracy and motion quality beyond joint acceleration, for example through command smoothness, contact stability, or task success? How much additional wall-clock latency does TTCO introduce in a realistic deployment setting?

Reviewer_kNVnOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents an inverse perspective on humanoid behavior foundation models: given a desired robot motion trajectory, the goal is to find the most suitable control commands to achieve it. The authors propose Test-Time Command Optimization (TTCO), which optimizes motion commands through sampling-based physics simulations at test time without retraining the BFM. The paper further investigates compute-optimal allocation strategies and demonstrates that TTCO can recover action labels from recorded robot trajectories, supporting downstream policy learning and demonstration reuse across different BFMs.

Critical strengths
  1. Clear motivation and practical significance. The paper addresses the command–execution gap in existing BFMs by optimizing motion commands at test time without retraining the underlying models. The motivation is clear.
  2. Practical value for demonstration reuse across BFMs. An interesting practical contribution is the ability to recover BFM-specific action labels from recorded robot trajectories, even when the original teleoperation commands are unavailable.
Critical weaknesses
  1. Improvements in tracking accuracy do not consistently translate into better downstream performance. Although TTCO substantially reduces motion-tracking errors, these improvements do not consistently translate into better downstream task performance. For example, in the SONIC → SONIC simulation setting, the average success rate drops from 75.8% with the original commands to 56.7% with the optimized labels. This suggests that improved tracking accuracy does not necessarily lead to better downstream policy performance, and the paper provides limited analysis of this discrepancy.

  2. Limited comparison with alternative target-specific adaptation methods. TTCO generates optimized action labels and trains a separate policy for each target BFM, whereas the original and state baselines reuse their existing labels and policies. This makes it difficult to isolate the benefits of TTCO itself from those of target-specific adaptation more generally.

  3. High computational cost and limited real-time applicability. The paper mainly measures computational cost in terms of simulated BFM transitions, while actual inference latency is also important for robotic applications. A more detailed analysis of the trade-off between computational cost and practical tracking improvements would strengthen the evaluation.

  4. Limited algorithmic novelty (minor). The algorithmic novelty is relatively limited, as TTCO primarily builds upon established MPC and sampling-based optimization techniques.

Questions for authors
  1. In the same-BFM simulation experiments, why do the optimized labels lead to substantially lower downstream task success rates than the original commands, despite the significant improvements in tracking accuracy? The paper demonstrates that TTCO can generate effective alternative action labels, but it has not sufficiently explained the performance gap between the restored labels and the original labels, nor has it identified the source of this gap.
  2. In the cross-BFM experiments, TTCO generates target-specific action labels, while the baselines directly reuse their existing labels. Could the authors further clarify the rationale behind this comparison?
  3. In Table 1, the reported body-position errors are very close to the root-position errors. To what extent are the large improvements in global tracking accuracy driven by root-position correction?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.