acceptodds
Withdrawn from ICLR 2027

Dynamic Humanoid Fall Recovery: A Dataset and Robust Recovery Policy

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Humanoid fall recovery remains challenging due to the highly diverse and out-of-distribution recovery states from real-world falls, and the conflict between rapid physical recovery and stable, natural motion. Existing approaches mainly address this challenge through multi-stage curriculum learning. However, such pipelines require complex multi-stage configurations and suffer from policy degradation due to distribution shifts induced by rigid stage transitions. To address these limitations, we propose KiPAR, a recovery-oriented reinforcement learning framework that combines dynamic state initialization, history-consistent motion priors, and phase-aware reward scheduling. KiPAR bypasses canonical static initializations by using a comprehensive dataset of diverse, dynamic dangerous states as initialization source. To ensure coherence after resetting to these highly unstable states, we introduce kinematic prior history injection to reconstruct physically consistent observation histories. Furthermore, we introduce a phase-aware reward scheme that dynamically balances heuristic incentives and stabilization regularizers. Extensive experiments demonstrate that KiPAR achieves a 98.34% recovery success rate, a 2.77s average time-to-stand, and a low 3.12% recollapse rate.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Humanoid Recovery, Reinforcement Learning, Dynamic Fall Dataset

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_JbxVOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This work investigates how to improve fall recovery for a physics-based humanoid controller. The authors propose gathering rollouts from a pre-trained general motion tracker during severe disturbances that meet a “failure” threshold. This dataset is then used during training to reset environments, and a phase-aware reward is used to balance recovery and tracking objectives. The authors show

Summarized contributions:

  • Fall Dataset
  • Kinematic Prior History Injection
  • Phase-aware reward scheme combining recovery incentives and reference tracking
Critical strengths
  • The idea of using a dataset of dangerous states is simple.

  • The phase-aware reward scheduling is intuitive.

Critical weaknesses
  • The submission has no supplementary video. This is crucial, as it makes it difficult to assess the dataset quality and the recovery policy. The screenshots are hard to judge because they are often miscropped and distorted.

  • It is unclear why a history reconstruction (Eq. (2)) is necessary if all the previous states are recorded exactly. See question 1.

  • The limitations section is insufficient and only mentions idealized sensing, which is not necessarily a limitation of the method itself. One limitation is the required pre-trained policy to collect rollouts.

  • The phase-aware reward is a very simple heuristic. Work on balancing rewards exists, e.g., [1, 2]; however, the authors do not sufficiently contextualize or investigate this contribution within the existing literature.

  • The Evaluation is minimal. The paper misses comparisons to any other recovery method. Evaluations contain ablations and comparisons against PHC.

Questions for authors
  1. The Fall Dataset contains the exact states (positions and velocities) of the look-back window. So why is a "history construction" even necessary when the actual velocities are collected?

  2. Typically, the reason the history is not properly reset is due to unknown quantities from the policy, such as past actions. This can be important information for the policy to learn regularization better, especially in real-world robot applications. How would your method deal with such a setup?

  3. The Fall Dataset is currently biased towards behavior resulting from a pre-trained controller. This is a) a costly process; you must first train a complete controller, and b) it limits general fall behavior. Can we not collect a dataset from random policies, or directly from poses?

  4. The phase coefficient for the reward formulation assumes a recovered state is at a certain height. If the dataset includes ground motions, we might also want to successfully track those; does this formulation still work?

Reviewer_D3cfOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses humanoid fall recovery from dynamic states instead of the usual static fall poses. The authors generate a fall dataset by pushing a PHC tracking policy around and labeling the frames before termination, propose KPHI, which back-extrapolates the current state to construct a short history at reset; and introduce a phase-aware reward scheme to balance heuristic incentives and stabilization regularizers. It achieves high success rate, short average time-to-stand and low recollapse rate in simulation. The ideas are worth investigating, but the paper needs external baselines, statistical characterization, complete metric definitions, and either real-world results or a substantially reframed scope.

Critical strengths
  • The motivation is solid. Rest-pose initialization really does throw away the velocity and contact history that matters when a robot is mid-fall, and KPHI is a sensible way to address this problem.
  • The phase-aware reward is well designed. Tying the heuristic annealing to root height and gating the regularization terms by a height/orientation band is a cleaner solution than a fixed hand-tuned schedule.
Critical weaknesses
  • The paper is framed around real-world deployment, but everything is in simulation. The abstract and intro talk about deployment being hindered, critical instabilities, and "extremely dangerous situations," yet every number comes from Isaac Gym with clean proprioception and contact. The limitations section concedes this in a single sentence, but the framing above it does not change. I also could not find an operational definition of "extremely dangerous situations" anywhere, which is overclaimed.
  • There is no real baseline. Table 1 compares against PHC and two variants of the authors' own recipe. HoST, HUMANUP, and HiFAR are all named in the related work as the state of the art in learned fall recovery, and none of them appear in the experiments. As it stands the paper only shows that the proposed reward combination beats its own ablations, which is not the same as showing it beats existing work.
Questions for authors
  • The experimental section lacks any external learned baseline. Table 1 only compares against PHC and variants of your own reward recipe, even though HoST, HUMANUP, and HiFAR are named in related work as the relevant prior methods. Could you add at least one of these under the same protocol? A reimplementation with published hyperparameters would already be much more informative than the current ablation-only comparison.
  • The metric definitions need work. "Upr." is never defined in the main text, and the re-fall metric appears as RCR in the text and caption, RFR in the setup, and RFT in §4.4. Energy, torque smoothness, and normalized jerk are reported to five significant figures with no units or definitions. Please define these in the main text, unify the naming, and clarify whether the success criterion in Table 1 is the same as the early-termination criterion in Appendix C.3.
  • The 2M+ to 26K subsampling is not explained. §4.1 says frames are filtered by uprightness, root height, and velocities, but gives no selection rule or target distribution. If this is stratified to over-represent hard states, it changes the training distribution and should be stated. Relatedly, Figure 5 is not linked to the easy/medium/hard bands in §4.1. Could you annotate the snapshots with their uprightness values so we can see what the hard 14.5% actually looks like?
  • Eq. S9 seems internally inconsistent. Root orientation and all velocities are held constant across the extrapolated history while position and joint configuration are linearly extrapolated. If the robot has nonzero angular velocity, orientation should be changing over that window; holding it fixed while extrapolating q is not the first-order approximation the text claims. It also makes the synthesized history less dynamic than the true pre-fall trajectory, which is the opposite of what KPHI is meant to preserve. Can you justify this, or show that it does not bias the synthesized history?
  • The KPHI gain in Table 2 is small. 98.34% vs. 96.67% for warm-up is a 1.67-point gap, with no confidence intervals, seed variation, or episode counts reported. Since the paper also concedes KPHI is slightly worse than warm-up on TTS and smoothness, I am not convinced the gap is meaningful. Please report variance and episode counts for Tables 1–3.
  • Figure 4(b) plots only root angular velocity; linear velocity is equally relevant to the "dynamic" claim, and the dataset stores it. Could you add root linear velocity?
  • A few remaining points: how sensitive is the method to history length k, which is a key KPHI hyperparameter but is not ablated? How do you justify 30 Hz policy / 60 Hz simulation relative to related humanoid recovery work? Does the low rate affect resolution of impact transients? What does "KiPAR" stand for, and can the main quantitative result be added to the abstract?
Reviewer_nhdWOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

KiPAR trains a simulated humanoid to recover from disturbed states collected during failed motion-tracking rollouts. Its main components are backward extrapolation to initialize observation history and rewards that change as the humanoid rises. The reported 98.34% recovery success rate is promising.

Critical strengths
  1. Dynamic initial states are a useful step beyond testing only fixed prone and supine poses.
  2. The history-initialization and reward-scheduling ablations are relevant. The appendix gives useful implementation and reward details.
Critical weaknesses
  1. The test split may overstate generalization. The authors randomly split individual frames from the same rollouts and motion clips (Section 4.1). Nearby frames are highly correlated, so test states may closely resemble training states. A split by rollout and AMASS motion, with results by difficulty and fall mode, is needed to support the generalization claim.
  2. The headline rate is dominated by easier states. The sampled dataset is 76.9% “easy” by uprightness and 14.5% “hard.” The paper does not report success separately for these groups. The overall 98.34% therefore does not establish reliable recovery from severe falls.
  3. The comparison does not establish an advantage over existing recovery methods. Table 1 compares KiPAR with PHC and variants of the authors’ reward setup, but not the fall-recovery methods discussed in Related Work. These internal comparisons support the value of the components, not a broader superiority claim.
  4. Stability and uncertainty are unclear. Success requires 60 consecutive upright steps and then terminates the episode during training. The paper does not clearly define how re-collapse is measured or whether evaluation continues after success. It also gives no run-to-run variation, confidence intervals, or exact evaluation episode count.
Questions for authors

Table 1 contradicts the statement that KiPAR has the lowest normalized jerk: the standing-reference variant reports a lower value. The authors should also specify whether time-to-stand, energy, and smoothness are averaged over all attempts or only successful ones; differing success rates make those comparisons sensitive to that choice. How does performance change under the sensor noise, delay, and physical variation used in training, given that evaluation uses nominal conditions?

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

Related papers

Discussion (0)

Sign in to comment.