acceptodds
Under review as a conference paper at ICLR 2027

HIPC: Humanoid Iterative Planning and Control via Goal-Conditioned Motion Closure

Abstract

Recent generative models are widely used to provide whole-body motion for humanoid tracking. Current methods combine motion generators with controllers in an open-loop paradigm, which leads to critical issues: tracking errors can accumulate into spatial drift, and the planner cannot adapt to unexpected physical disturbances without feedback from physical execution. However, directly feeding realized states back to the planner in a closed loop can cause discontinuities, resulting in physical jitter and even falls, since such states are out of distribution during training. To bridge this gap, we propose HIPC, a real-time Humanoid Iterative Planning and Control framework that couples a generative model with motion tracking into a unified planning-and-control paradigm. HIPC introduces a goal-conditioned motion planning trained with a canonical motion representation that avoids gimbal-lock singularities under extreme orientations and supports motion generation on diverse geometric and semantic conditions. Then, HIPC adopts two strategies, deviation-decayed recovery and history re-anchoring, to mitigate the out-of-distribution shift caused by realized-state feedback. Finally, HIPC operates a real-time asynchronous planning-and-control system, with latency compensation to reduce temporal mismatch for inference delay. Extensive experiments confirm the effectiveness of HIPC under switching and composition of diverse conditions. Results show that HIPC achieves state-of-the-art goal-reaching accuracy among existing methods while simultaneously demonstrating robustness to physical perturbations and the ability to recover from falls.

Primary area:
applications to robotics, autonomy, planning
Keywords:
motion generation, humanoid control

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_zezKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents incremental work on motion generation for humanoid locomotion. The starting point is a two thread detached system with one planner (the proposed HIPC) and a low level tracker/controller (SONIC, Luo et al, 2026). This two thread setup is similar to many earlier methods. In the proposed setup these are detached and as the generation of a plan takes some time, and since actions may not be perfectly executed, the plan will always be made with an incorrect initial state, and it will also be slightly out of date when it reaches the controller. The paper describes two strategies to fix these problems, these are called: deviation delayed recovery, and history re-anchoring.

Critical strengths
  • Evaluation uses mostly metrics used in the cited papers e.g. (Xie, 2026) and (Chen, 2026), but also introduces some of its own in order to test Q1,Q2,Q3.
  • The experiments demonstrate the effectiveness of the proposed modifications.
Critical weaknesses
  • Although the proposed modifications make sense, I remain unconvinced that more closed-loop control like ReactiveBFM is not a better approach. E.g. Chen et al. 2026 describes approaches using a SONIC style tracker as "severely crippling its dynamic reactivity to physical feedback". I believe this weakness remains, even if the two proposed fixes are applied. Even worse, ReactiveBFM by Chen et al. 2006 is here coupled with the SONIC tracker (311), exactly the approach they complain about!

  • The details on how the datasets are used is not entirely clear. It seems as if datasets AMAS, LAFAN, HiPHI and GMR are used for training (237-239) and the entire 100STYLE for testing (not just the test set asin Chen, 2026) (361,400), together with Replica indoor scenes (410). Is this correct?

  • How were the other methods in the result tables trained? Were they retrained on the same datasets as HIPC, if so how? Or did you use their weights? If so, how about data leakage from 100STYLE?

  • The paper is quite hard to follow, due to numerous easy to fix problems with the text. E.g.

    • #183 Why are horizontal quantities incremental, end vertical states direct? This is detail just hanging in the air.
    • #189 "where R_t in SO(3) and W^p_t in R^3". These variables are NOT in the Eq. (2) above. Better to introduce them before Eq. 2, and describe the variables in Eq. (2) below Eq.2.
  • The related work discussion is mostly good. Missing is the OMG paper that is compared to in the results tables.

  • The result tables list methods ARDY and OMG. These are not properly cited. OMG is by S. Huang et al. (2026) and should be cited and discussed in the related work, ARDY is cited, but not referred to as ARDY in the citation nor in the related work, making the connection not obvious.

Questions for authors

Q1: Can you provide performance for ReactiveBFM and OMG using their complete systems, instead of always coupling them with SONIC (that ReactiveBFM explicitly states as a bad idea)?

Q2: Can you detail how HIPC and the other compared methods were trained? Datasets, use of public weights etc.?

Q3: The looser coupling compared to ReactiveBFM is potentially reducing the compute load. Is this the case, and if so, can it be demonstrated somehow?

2 comments

We argree system comparison should be more clear, we present more experiment details as follows:

Weakness 1 & Q1: Reactivity and Full-System Comparisons

Planner–tracker coupling. We agree that universal trackers alone cannot adapt high-level motion plans to changing instructions. However, ReactiveBFM's criticism concerns this reliance on predefined references and SONIC's keyframe-in-betweening planning. HIPC addresses this limitation by replacing SONIC's original planner with a goal-conditioned motion generator. ReactiveBFM likewise employs a separately pretrained ScaleBFM tracker instead of jointly training its planner and controller. ReactiveBFM does not test its planner with SONIC, so the cited criticism does not establish that SONIC is unsuitable for closed-loop motion control.

ReactiveBFM comparison. We reproduce ReactiveBFM's Gaussian prefix-noise augmentation within our training framework. For each prefix frame t and motion dimension d, we independently sample

\quad \rho=\operatorname{clip}(s/S,0,1),$$ where s/S is the training progress and $\sigma_d$ is the dataset standard deviation of the d-th unnormalized motion dimension. Noise is added to the normalized motion prefix. We train a **ReactiveBFM baseline** using this augmentation. Both checkpoints share the HIPC planner architecture and training data, **the only training difference is Gaussian prefix noise versus deviation-decayed recovery**. At inference, we compare history construction and asynchronous execution strategies under the same SONIC tracker and 2 Hz replanning frequency. Full details are provided in Table 3. | Framework | History | Async. Strategy | E_pos (m) ↓ | Fall (%) ↓ | |---|---|---|---:|---:| | ReactiveBFM | Realized | Direct Switching | 1.386 | 85.714 | | ReactiveBFM | Realized | Temporal Ensembling | 1.397 | 76.190 | | ReactiveBFM | Hybrid | Temporal Ensembling | 0.089 | **0.000** | | ReactiveBFM | Hybrid | Latency Compensation | 0.076 | **0.000** | | HIPC | Hybrid | Direct Switching | 0.184 | 1.190 | | HIPC | Hybrid | Temporal Ensembling | 0.097 | **0.000** | | **HIPC** | **Hybrid** | **Latency Compensation** | **0.067** | **0.000** | **Results.** Realized histories produce large errors and frequent falls despite prefix-noise training, suggesting that independent Gaussian noise does not adequately capture execution-induced deviations. Hybrid histories combined with temporal ensembling or latency compensation achieve zero falls for both checkpoints, with latency compensation further improving accuracy. This is a **controlled mechanism comparison**, not a reproduction of the original ReactiveBFM–ScaleBFM system. ## Weaknesses 2–3 & Q2: Datasets, Weights, and Leakage HIPC trains on **AMASS, LAFAN, and HiPHI**, retargeted through GMR. The original evaluation used 100STYLE as test set, and **Replica** scenes served separate dynamic tracking tests. **Kimodo, ARDY, TextOp, and OMG use public weights without fine-tuning.** We additionally evaluate OMG's **native generator–HoloMotion tracker** (OMG\*). All use SONIC except OMG\*. To prevent data leakage with 100STYLE in OMG's training data, we re-evaluate on **MotionDecode**, which is not listed among the compared methods' reported training datasets. Selected results are shown below, with complete comparisons in the revised paper. Full details are provided in Table 2. | Method | PJ → | AUJ ↓ | Fall (%) ↓ | E_pos (m) ↓ | FID ↓ | |---|---:|---:|---:|---:|---:| | OMG | 1.109 | 7.327 | 22.000 | -- | 0.592 | | OMG\* | 0.522 | 4.595 | 1.333 | -- | **0.533** | | **HIPC** | 0.445 | **2.215** | **0.000** | **0.077** | 0.592 | ## Q3: Computational Cost and Coupling **Response:** ReactiveBFM, like HIPC, employs a separately pretrained, frozen tracker (ScaleBFM). Both systems incorporate execution feedback at inference time, differing primarily in **how realized states are incorporated into planner histories**. A more tightly coupled approach is Humanoid-DART, which uses simulated tracker executions to iteratively improve both the planner and tracker during training. HIPC instead focuses on mitigating execution-induced distribution shift through history re-anchoring, without controller-specific training. We claim no training or inference efficiency advantage over ReactiveBFM. If the reviewer considers ReactiveBFM more tightly coupled or more closed-loop, we would appreciate clarification of the specific mechanism underlying this distinction. ## Presentation and Citations We explain that incremental horizontal quantities ensure planar position/heading invariance, while direct vertical states preserve ground-relative height/posture (191); $R_t$ and ${}^W p_t$ now precede Eq. (2) (195). We added OMG (Huang et al., 2026) to Related Work (144), identify ARDY by name, and corrected both experimental citations (362).

Dear Reviewer,

Thank you again for your valuable comments. As the discussion period is nearing its end, we would like to kindly ask whether our responses have adequately addressed your concerns.

Please let us know if any further clarification is needed. If our responses have resolved your concerns, we would greatly appreciate it if you could reconsider your evaluation in light of the additional evidence and clarifications provided.

Thank you for your time and consideration.

Best regards,
The Authors

Reviewer_pREaOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper developes a system that is an integration of a path planner (A* or OmniNav, preexisting), a generative motion planner that takes those waypoints (novel), and a preexisitng whole body controller called SONIC. Their main contribution, the mid-level planner runs at 2Hz while the lower level controller at 50Hz. The authors claim that their main contributions are (1) motion representation and conditioning (questionable novelty), (2) execution feedback (history reanchoring with measured state to handle tracking error) (again questionable as this sounds like MPC) and (3) latency compensation. To validate this, the authors do simulations and some experiments with G1 robot and do comparison with motion generation baselines like Kimodo, ARDY, TextOp, and OMG, ablation studies to tease out the value of their proposed blocks. I understand that the results on a high level show 0.10 m position error, zero observed falls in selected evaluations, and 53% recovery success after induced falls. My overall impression of this paper is that it provides an interface between a learned motion generator and an existing tracker, rather than a new low-level controller or a new principle of hierarchical planning.

Critical strengths

The paper targets a practical integration problem of modules operating at different rates with latencies and changing scenes. In terms of presentation, Figures 1,2,3 earn their place in this paper, they are well sketched and convey the ideas of the paper nicely. Especially the figure 3 demonstrating latency compensation is information dense and conveys quickly what their algorithm does. The paper also does comparisons with many baseline reference planners so it is very comprehensive on that score. The ablations are informative as the experiments investigate the value of history construction, latency compensation and recovery augmentation teasing out which components improve performance.

Critical weaknesses

The authors claim that their novelty lies in integrating a generative motion planner with an underlying tracking controller (from other works) and then fixing issues like jittery motion, inference latency compensation that arise from stitching the generative planner with the controller. In several places, the title of the paper is incorrect; in some places they say “iterative planner” while in others they say “interactive planner”, the authors should correct that. In my assessment, this paper does not offer any new science or insight that isn’t already known and implemented exhaustively in production robotics stacks. For example, the paper claims their motion representation being singularity free as a contribution but it is at best modest because gravity-vector plus relative-rotation representations are standard in legged RL. The authors should cite references to their claims so we know which years those references are from. I think global representations using rotation matrices and quaternions are widespread now. The ideas of separating slow planning from fast feedback control, doing replanning from measures states, expressing goals in robot’s base frame, and accounting for inference delay (which is the same as computation delay) are well known design principles that are abound in self-driving cars, warehouse robots and cobots. Giving these principles new names like “motion closure”, “fast-in-slow dual-system architecture”, “execution distribution gap” does not itself establish novelty. The authors should consider describing how their motion closure approach relates to receding horizon planning like MPC, if there are any similarities or differences (except from model-free and model-based ofcourse). The asynchronous separation of planning and control, and compensation for planning latency, are established robotics practices (think delay-compensated MPC or Smith-predictor-style designs). The paper should clarify what is technically novel about its predictive history-shifting mechanism relative to existing approaches like MPC, and demonstrate its benefit against relevant latency-compensation baselines. Improvements over uncompensated trajectory switching alone do not establish that novelty. ReactiveBFM is a close prior work and appears in related work and Table 1 but is not evaluated or compared against in the experiments. I am also concerned that their planner has no environment awareness as it does not consume the occupancy map, and navigation success of 70% (OmniNav) and 90% (A*) mostly reflects the upstream planners.

Questions for authors

None

4 comments

We agree that our terminology and contributions should be clarified, but respectfully disagree that the paper offers no novel contributions. To the best of our knowledge, no existing method jointly enables precise reaching of fine-grained geometric goals while following semantic instructions, together with robust recovery from physical perturbations. We clarify our contributions below.

W1: Framework Name

Response: We have standardized iterative in the paper.

W2: Novelty, Motivation, and MPC

Response: We do not claim novelty in fast-in-slow dual-system architecture, or execution distribution gap. Fast-in-slow relates to RoboDual (2024), GR00T N1 (2025), and Fast-in-Slow (2025); distribution shift to DAgger (2011) and ReactiveBFM (2026).

In this paper, our focus is precise, robust spatial goal reaching under imperfect execution. HIPC additionally quantifies continuous geometric errors beyond goal-reaching success rates. We address three challenges:

  1. Motion representation and feedback. To achieve precise goal reaching and robustness to disturbances, we design a canonical motion representation that supports arbitrary body orientations including fallen states. We establish two complementary feedback mechanisms: goal re-expression updates spatial goals relative to the current robot state to correct accumulated drift, while history feedback incorporates realized states into subsequent motion generation.
  2. Handling execution-induced distribution shift. Directly feeding controller-executed histories into a generative planner can introduce out-of-distribution motion patterns. We instead re-anchor generated histories to the latest realized state, preserving their relative motion structure. Deviation-decayed recovery further trains the planner to handle such deviations and recover toward plausible motions.
  3. Asynchronous execution and latency compensation. During planner inference, the controller continues executing the previous reference, causing the new plan to become outdated before execution. Building on history re-anchoring, we shift the predicted history to the anticipated handoff time and re-express the goal accordingly, aligning both history and goal conditions without additional dynamics prediction.

Relationship to MPC. HIPC shares receding-horizon feedback with MPC but employs a learned generative motion planner instead of explicit trajectory optimization, supporting diverse semantic and geometric conditions and their combinations (additional results in Supplementary Sec. 1). Our focus is maintaining precise goal tracking and satisfying semantic instructions under physical execution deviations and perturbations.

W3: Latency Compensation

Response: We agree that delay compensation is established in robotics. HIPC's predictive history shift builds directly upon history re-anchoring (Eq. 3) (319): predicted histories provide future motion references but lack execution feedback, whereas realized histories incorporate feedback but lack future states. Our hybrid history combines both, reusing existing generated kinematics to shift the re-anchored history and re-express spatial goals at the anticipated reference handoff, without an additional dynamics predictor. Rather than introducing a separate prediction mechanism, HIPC addresses temporal misalignment through the same hybrid history construction used for execution feedback.

W4: ReactiveBFM Comparison

Response: We add a controlled ReactiveBFM comparison. Full details are provided in Table 3.

FrameworkHistoryAsync. StrategyE_pos (m) ↓Fall (%) ↓
ReactiveBFMRealizedDirect Switching1.38685.714
ReactiveBFMRealizedTemporal Ensembling1.39776.190
ReactiveBFMHybridTemporal Ensembling0.0890.000
ReactiveBFMHybridLatency Compensation0.0760.000
HIPCHybridDirect Switching0.1841.190
HIPCHybridTemporal Ensembling0.0970.000
HIPCHybridLatency Compensation0.0670.000

W5: Navigation Ablation

Response: To isolate HIPC's contribution to navigation, we added an ablation on 50 tasks per navigation method, with and without external forces. The open-loop variant does not incorporate realized-state feedback into the HIPC planner.

NavigationConditionOpen-loopClosed-loop
A*No perturbation29/5045/50
Force perturbation24/5045/50
OmniNavNo perturbation25/5035/50
Force perturbation22/5034/50

Success thresholds: 0.5 m (A*) and 2 m (OmniNav).

HIPC improves navigation success even with OmniNav's closed-loop visual feedback, demonstrating additional spatial drift correction. Its nearly unchanged success rates under external forces further demonstrate perturbation robustness. Full details are provided in Table 4.

Dear Reviewer,

Thank you again for your valuable comments. As the discussion period is nearing its end, we would like to kindly ask whether our responses have adequately addressed your concerns.

Please let us know if any further clarification is needed. If our responses have resolved your concerns, we would greatly appreciate it if you could reconsider your evaluation in light of the additional evidence and clarifications provided.

Thank you for your time and consideration.

Best regards,
The Authors

Thanks for the detailed response and the additional experiments. I appreciate the effort the authors have put into addressing my concerns.

The new comparison with ReactiveBFM and the navigation ablation are helpful. In particular, the open-loop vs. closed-loop comparison provides better evidence that the feedback mechanism improves navigation performance, rather than the improvement coming entirely from the upstream planner. The clarification of the history re-anchoring and latency compensation mechanisms is also useful.

However, my main concern is still about the novelty of the proposed method. I agree that integrating a generative motion planner with a tracking controller and making the system work reliably is a useful engineering contribution. But most of the underlying ideas, including receding-horizon replanning, execution feedback, history alignment, and latency compensation, are already well established in robotics.

The new ReactiveBFM comparison is interesting. Once hybrid history and latency compensation are applied to both methods, their performance becomes quite similar (0.076 m vs. 0.067 m position error, both with zero observed falls). To me, this suggests that the main benefit comes from the history construction and execution strategies, rather than a fundamentally different planning approach. While these are useful design choices, I am still not fully convinced that they represent a sufficiently significant methodological contribution.

I also appreciate the clarification of the relationship to MPC. However, using a learned generative planner instead of explicit trajectory optimization does not, by itself, establish the novelty of the closed-loop planning mechanism.

Overall, I think the rebuttal has strengthened the experimental evaluation and clarified the contributions. I see the value of the proposed system, especially for practical humanoid motion generation and execution. However, my concern about the limited methodological novelty remains, so I will keep my current rating.

We appreciate the reviewer's thoughtful follow-up. We would like to clarify that our contribution is not the invention of a new closed-loop system, but a fine-grained geometric closed-loop mechanism that jointly constrains the generator and controller, enabling: (1) Precise goal execution; (2) Robust state recovery; and (3) Coupled control under diverse conditions, including compositional geometric and text conditions.

Achieving fine-grained state control at the body level, with diverse, precise, and robust autonomous motion expression, has long been a goal of humanoid robotics. While ReactiveBFM supports text instructions and target positions, root-only geometric goals are insufficient to specify diverse whole-body configurations, leaving fine-grained state control at the body level an unresolved challenge.

Again, we emphasize that humanoid robots have important properties, including high degrees of freedom and underactuated control in dynamics. Therefore, introducing established concepts such as closed-loop feedback, replanning, and latency compensation into humanoid robotics remains challenging. This motivates us to design a unified planning-and-control framework that learns a multi-conditioned motion distribution to autonomously release the diversity and fine granularity of humanoid body expression.

To address these challenges, we jointly design geometric goal conditioning, training strategies, and the planner–controller interface to adapt established planning and control principles to high-DoF humanoid motion generation and tracking. Specifically, history re-anchoring is tailored to the high-DoF humanoid state, applying rigid-body transformations to root poses and additive deviations to joint configurations. This state construction also underlies deviation-decayed recovery and latency compensation via history shift, forming a unified mechanism for execution feedback, recovery, and asynchronous planning. Furthermore, our motion-clip construction and sampling strategies explicitly consider geometric goal coverage and fallen-state recovery during training. Together, these designs enable HIPC to support multimodal geometric and semantic conditions, independently or in combination, achieving precise goal reaching and guided fall recovery from real state within a unified planner.

We respectfully disagree that HIPC lacks methodological novelty. Our contribution is to address the challenge of body-level fine-grained state control in high-DoF humanoids, bridging the gap between expressive motion generation and precise physical execution. The methodological novelty lies in the joint design of geometric goal conditioning, history re-anchoring, deviation-decayed recovery, and latency-aware history shifting to tackle this challenge. Together, these mechanisms enable diverse, precise, and robust control under compositional geometric and text conditions. Extensive experiments validate these capabilities through accurate goal execution, robust state recovery, and multimodal compositional control (particularly Table 5).

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

This paper presents HIPC, a system that combines a goal-conditioned motion planner with a frozen humanoid motion tracker. Its main idea is to preserve the structure of the planned motion history while aligning it with the robot’s current state. The planner is trained with perturbed histories and recovery trajectories whose deviations gradually decay, and a predictive history shift accounts for inference latency. Experiments on the Unitree G1 cover goal reaching, external disturbances, fall recovery, and real-world execution. The reported results show improved goal accuracy over the evaluated baselines, while internal ablations support the proposed history construction and recovery training.

Critical strengths
  1. The paper offers a practical way to connect motion generation with execution feedback. Re-anchoring a planned history to the current robot state is a simple and useful design. Together with recovery training and latency compensation, it addresses concrete problems in deploying generative planners with an existing tracker.
  2. The evaluation includes meaningful system-level evidence. The experiments cover several goal types, disturbances, and recovery from fallen states, with real-world demonstrations on G1. Tables 3 and 5 also provide useful evidence that the history construction and recovery training matter within the proposed system.
Critical weaknesses

The benefit over closely related closed-loop methods remains unclear. ReactiveBFM already uses execution feedback, self-rollout and prefix-noise training, and asynchronous planning with temporal ensembling. HIPC’s re-anchored history, recovery targets, and predictive history shift offer distinct designs, but ReactiveBFM is not included in the quantitative comparisons. Table 3 provides useful internal evidence, although it does not specify whether the history variants were trained with corresponding input constructions. This limits what the results establish about the advantage over a planner trained for realized-history feedback.

Questions for authors
  1. History variants and comparison with ReactiveBFM. Were the predicted-, realized-, and hybrid-history variants in Table 3 trained separately with corresponding history constructions, or evaluated by changing the inputs to the same checkpoint? To assess the benefit over ReactiveBFM’s approach, could you provide a matched comparison on shared tasks, or a controlled realized-history baseline with self-rollout and prefix-noise training, using the same tracker and replanning frequency?
  2. Recovery augmentation. In Table 5(b), does removing deviation-decayed recovery preserve the number and sampling frequency of get-up examples? Please specify the perturbation ranges, decay function, and recovery-horizon distribution. A comparison with prefix-only perturbation while holding get-up sampling fixed would help determine whether the decaying future targets provide an additional benefit.
2 comments

We appreciate your constructive responses, here we aim to address all your questions and concern below:

Weakness & Q1: History Ablation and ReactiveBFM Comparison

Original Table 3. The predicted-, realized-, and hybrid-history variants use the same HIPC checkpoint, differing only in inference-time history construction. No controller-executed trajectories enter planner training. ReactiveBFM does not report training its planner on trajectories executed by its downstream ScaleBFM tracker; in both methods, realized-state feedback is incorporated at inference time.

Controlled comparison. We reproduce ReactiveBFM's Gaussian prefix-noise augmentation within our training framework. For each prefix frame t and motion dimension d, we independently sample

\quad \rho=\operatorname{clip}(s/S,0,1),$$ where s/S is the training progress and $\sigma_d$ is the dataset standard deviation of the d-th unnormalized motion dimension. Noise is added to the normalized motion prefix. We train a **ReactiveBFM baseline** using this augmentation. Both checkpoints share the HIPC planner architecture and training data, **the only training difference is Gaussian prefix noise versus deviation-decayed recovery (DDR)**. At inference, we compare history construction and asynchronous execution strategies under the same SONIC tracker and 2 Hz replanning frequency. We also replace 100STYLE with **MotionDecode** to avoid potential data-leakage. | Framework | History | Async. Strategy | E_pos (m) ↓ | Fall (%) ↓ | |---|---|---|---:|---:| | ReactiveBFM | Realized | Direct Switching | 1.386 | 85.714 | | ReactiveBFM | Realized | Temporal Ensembling | 1.397 | 76.190 | | ReactiveBFM | Hybrid | Temporal Ensembling | 0.089 | **0.000** | | ReactiveBFM | Hybrid | Latency Compensation | 0.076 | **0.000** | | HIPC | Hybrid | Direct Switching | 0.184 | 1.190 | | HIPC | Hybrid | Temporal Ensembling | 0.097 | **0.000** | | **HIPC** | **Hybrid** | **Latency Compensation** | **0.067** | **0.000** | **Results.** Despite Gaussian prefix-noise training, realized histories yield substantial errors and frequent falls, suggesting that independent Gaussian noise alone does not adequately capture execution-induced deviations. In contrast, **hybrid histories preserve generated motion structure while incorporating execution feedback**, substantially improving accuracy and stability. Combined with temporal ensembling or latency compensation, hybrid histories achieve zero falls for both training variants. Latency compensation further reduces position error in both cases, with HIPC achieving the lowest error of **0.067 m**. These results demonstrate the consistent benefits of **history re-anchoring and latency compensation** across different training augmentations. ## Q2: Controlled Deviation-Decayed Recovery Ablation To isolate the contribution of future deviation decay from increased get-up sampling, we train a **prefix-only deviation baseline** with identical get-up sampling and training settings. It applies deviations only to the history, leaving the future ground-truth trajectory unchanged. Full details are provided in Table 6(b). | Variant | Recovery (%) ↑ | T_getup (s) ↓ | |---|---:|---:| | DDR w/o Future Decay | 13.333 | 6.745 | | **HIPC** | **53.333** | 7.227 | With identical get-up sampling, future deviation decay improves recovery success from **13.33% to 53.33%**. Without future decay, the deviation window ends abruptly, introducing discontinuities in the training trajectories. This controlled ablation demonstrates that **smooth recovery supervision provides benefits beyond increased get-up sampling alone**. The perturbation ranges and recovery horizons are summarized below. Root rotations are defined in local axes (x-forward, y-left, z-up); U[a,b] denotes a uniform distribution, U_Z denotes a discrete uniform distribution, and joint offsets are constrained by joint limits. Full details are provided in Table 8 of Appendix. | Perturbation | Online Training | Offline Get-Up | |---|---|---| | **Root state** | | | | Local x rotation (rad) | U[-0.20,0.20] | U[-0.05,0.05] | | Local y rotation (rad) | U[-0.20,0.20] | U[-0.10,0.10] | | ... | ... | ... | | **Joint configuration (rad)** | | | | Shoulder | U[-0.30,0.30] | U[-1.50,1.50] | | Elbow | U[-0.10,0.10] | U[-1.50,1.50] | | ... | ... | ... | | **Recovery schedule** | | | | Horizon (frames) | U_Z[15,40] | U_Z[20,50] | | Duration (s) | 0.30–0.80 | 0.40–1.00 | Both online and offline augmentations use a quintic decay function: $$w(u)=1-6u^5+15u^4-10u^3,\quad u=\operatorname{clip}(k/K,0,1),$$ where k is the frame offset and K is the sampled recovery horizon. Joint and root-height deviations are scaled by w, and orientation deviations use SLERP. The decay has zero first- and second-order derivatives at both endpoints, ensuring smooth transitions without additional velocity or acceleration discontinuities.

Dear Reviewer,

Thank you again for your valuable comments. As the discussion period is nearing its end, we would like to kindly ask whether our responses have adequately addressed your concerns.

Please let us know if any further clarification is needed. If our responses have resolved your concerns, we would greatly appreciate it if you could reconsider your evaluation in light of the additional evidence and clarifications provided.

Thank you for your time and consideration.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.