acceptodds
Under review as a conference paper at ICLR 2027

StableBFM: Extending Behavior Foundation Models to Robust Behavior

Abstract

Pretrained behavior foundation models (BFMs) provide a reusable foundation for whole-body humanoid control by conditioning on large-scale reference motions and decoding actions, enabling high-fidelity tracking of diverse behaviors. However, real-world deployment also requires reliable closed-loop recovery from off-nominal states such as falls. How to extend the capabilities of a pretrained BFM to recovery while preserving its existing tracking ability and keeping training overhead low remains a key challenge. Given the current robot state and a reference motion, we define the state–reference discrepancy as the deviation between the current state and the target state implied by the reference. Standard tracking pretraining mainly constrains behavior in the small-discrepancy regime, whereas fall recovery requires the decoder to start from large-discrepancy states, reduce the discrepancy through closed-loop control, and return to the nominal tracking region. From this perspective, we formulate recovery as extending the effective closed-loop control range of a pretrained BFM, and unify tracking and recovery as the same state–reference conditional decoding problem across discrepancy regimes. To this end, we treat recovery as a capability extension of a pretrained motion-tracking BFM and perform simple, low-overhead post-training without adding model parameters. We retain the original tracking PPO objective, allocate a subset of parallel environments to recovery, and apply DAgger-style expert supervision to states visited by the learner, thereby extending recovery supervision to the large-discrepancy regime. We further design corresponding recovery-injection mechanisms for two decoder parameterizations. Experiments show that the proposed framework significantly improves recovery while maintaining tracking performance, and supports deployment on a real humanoid robot.

Primary area:
applications to robotics, autonomy, planning
Keywords:
reinforce learning, behavior cloning, humanoid control, motion tracking, fall recovery

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_S6rFOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how to add fall recovery to a pretrained humanoid motion tracker while retaining a single deployed decoder and an unchanged state–reference interface. StableBFM continues PPO training on tracking environments and uses a recovery expert to label learner-visited states in a separate recovery branch. Both objectives update the same decoder. The paper considers an MLP and a context-conditioned Transformer, studies the frequency of recovery updates, and introduces cache perturbations for recovery supervision in the latter. Evaluation uses held-out AMASS and Motion Decode motions, a 192-trial fall-to-stand benchmark, and pushed tracking rollouts in MuJoCo after training in Isaac Lab. A Unitree G1 sequence provides qualitative hardware evidence. The MLP variant reports 93.23% eventual stand-up success and 73.60% motion recovery, with a reported post-training cost of two GPU-hours.

Critical strengths
  1. The deployment constraint is useful and the recovery gains are substantial. The recovery expert is used during training, while inference retains the existing action decoder and control interface. Table 2 shows that the MLP acquires both stand-up recovery and return to an advancing reference, whereas the released tracker has zero recovery under these criteria.

  2. The evaluation distinguishes several behaviors that are often conflated. The paper separates knockdown resistance, root recovery, upright recovery, motion recovery, and clean versus retried stand-up outcomes, and supplies numerical thresholds in Appendix B. The held-out motion sets and the update-frequency/context ablations are valuable. For example, the best setting for eventual stand-up success differs from the best setting for motion recovery (Table 3), which is an informative practical observation.

Critical weaknesses
  1. The evidence supports a tracking–recovery trade-off more clearly than preservation of tracking performance. On Motion Decode, the MLP's tracking success decreases from 82.38% to 76.85%, while global MPJPE increases from 547.30 to 1181.90 mm (Table 2(a), p. 8). The context model also drops from its own reproduced baseline's 79.51% to 71.32% on this dataset. Near-preserved AMASS success does not resolve these losses. The harmonic Stability Score combines capabilities but does not establish that the original capability has been retained to an acceptable degree. This matters because preserving a pretrained tracker's behavior is a central motivation, rather than a secondary convenience.

  2. The comparison does not yet fully separate recovery capability, evaluation selection, and training resources. Appendix B conditions recovery on each model's actual knockdowns, although knockdown rates differ; consequently the models need not be evaluated on the same subset of disturbed episodes for this metric. Appendix E also gives the STUBBORN reproduction rough-terrain training while the other recovery runs use flat ground. The two-GPU-hour figure is a post-training measurement, and the cost of training the reproduced BFM-Zero expert is not reported in that accounting. These choices do not invalidate the method, but limit the interpretation of direct superiority and efficiency claims. Training-seed variability is not reported, and the hardware evidence in Section 4.5 is qualitative, leaving the reliability of the complete trade-off uncertain.

Questions for authors
  1. What level of tracking degradation is considered acceptable for the claimed preservation of the pretrained model? Can you explain the Motion Decode losses in Table 2, including the substantial increase in global MPJPE, and indicate whether a checkpoint or training setting yields a better tracking–recovery trade-off?

  2. How many motions remain in the common no-push scored set, and how many pushed trials contribute to each model's conditional recovery denominator? Please report successful continuation over all matched pushed trials, or recovery over an explicitly shared knockdown subset, alongside the existing conditional rates.

  3. What is the cost of obtaining the recovery expert, and how does the comparison change when that cost is included or explicitly treated as an amortized prerequisite? How sensitive is the STUBBORN comparison to matching the terrain and post-training conditions?

5 comments

Q1. Tracking preservation, Motion Decode errors, and training configurations

(a) Tracking preservation

For joint motion tracking and recovery, the MLP’s 5.53-point change in Motion Decode success remains acceptable. The model retains 76.85% success while achieving 93.23% eventual stand-up success and 73.60% motion recovery. Table 2 shows good preservation of motion quality: local MPJPE increases by only 10.47 mm (53.35 → 63.82 mm), PA-aligned error by 5.43 mm (32.82 → 38.25 mm), and root-height error decreases from 60.51 mm to 58.02 mm. On AMASS, success (82.80%) and local MPJPE (52.64 → 52.43 mm) are nearly unchanged. The pretrained controller’s coordination remains intact.

(b) Global MPJPE growth: dataset amplification

We explain the larger global-error increase on Motion Decode by its sensitivity to accumulated global errors. For the same MLP, MPJPE-G increases by 76.10 mm on AMASS (606.50 → 682.60 mm) and 634.60 mm on MD (547.30 → 1181.90 mm), while MD local and PA errors increase by only 10.47 mm and 5.43 mm. Small persistent velocity or heading deviations accumulate into trajectory-position errors, and heading-induced position errors also depend on motion displacement. Clip duration and motion distribution therefore affect how the same global tracking error appears in the metric. We tested this by injecting identical perturbations into ideal reference trajectories from Motion Decode (1503 clips) and AMASS (1500 clips reconstructed with the appendix’s stratified sampling protocol, seed 20260902).

Table D. Dataset characteristics and sensitivity to global errors.

Data characteristic / injected errorReported metric (unit)Motion Decode (1503 clips)AMASS (1500 clips)
Clip durationMedian (s)23.16.7
Maximum displacement from the starting pointMedian (m)0.410.71
Displacement from the starting point exceeding 3 mFraction of clips (%)19.25.1
Displacement from the starting point exceeding 5 mFraction of clips (%)13.30.9
Displacement from the starting point exceeding 10 mFraction of clips (%)6.10.0
Cumulative turningMedian (°)305125
Heading drift of 1°/sstrict-foot SR(%)5581
Random walk, σ=0.05 m/√sstrict-foot SR(%)1249
Fixed heading offset of 3°strict-foot SR(%)8498
Stride-length error of 5%strict-foot SR(%)8498
Fixed heading offset of 10°strict-foot SR(%)6855
Heading drift of 1°/sMPJPE-G increase (mm)+293+46

SR denotes strict-foot diagnostic success after data-level error injection, using a 0.20 m world-coordinate threshold between corresponding feet. Table 2 uses the original no-foot evaluation. MPJPE-G increases are computed relative to the ideal trajectories before error injection.

Table D shows the amplification: MD has longer clips (23.1 s vs. 6.7 s) and a heavier tail of long-distance motions. Under identical heading drift of 1°/s, MPJPE-G increases by 293 mm on MD and 46 mm on AMASS. Thus, the same global tracking change incurs a larger metric cost on MD. Meanwhile, local/PA errors increase by 10.47 / 5.43 mm for the MLP and 20.90 / 12.38 mm for the Transformer. Local execution remains strong, while global deviations are amplified more on MD.

(c) Configuration selection

Our hyperparameters were selected through Pareto-front tradeoff analysis. In Table 3(a), K=6 trades only 0.33 points of AMASS tracking success relative to K=2 for an 8.80-point gain in motion recovery (73.60% vs. 64.80%). Compared with K=3 (69.80%), K=6 better serves the goal of reconnecting with the advancing reference.

We ablated the recovery-environment fraction ρ, behavior-cloning weight β_rec, and initial-pose pool size (Table R).

Table R. Single-parameter ablations on Motion Decode.

ConfigurationRecovery environments / total environmentsInitial posesBC coefficientTracking success (%)Local / global MPJPE (mm)Eventual stand-up success (%)Motion recovery (%)
Default1024 / 4096353276.8563.82 / 1181.9093.2373.60
Fewer recovery environments512 / 4096353276.2164.35 / 1194.7291.4871.12
More recovery environments2048 / 4096353276.6963.98 / 1186.0293.5273.92
Expanded initial-pose pool1024 / 4096500276.9163.74 / 1180.2793.3573.79
Lower BC coefficient1024 / 4096353176.5864.08 / 1188.4191.7270.94
Higher BC coefficient1024 / 4096353475.9665.21 / 1206.8593.4773.18

Two findings follow: (1) the default is already balanced; weakening recovery supervision reduces both recovery and tracking success, and increasing β_rec to 4 lowers tracking success to 75.96%. (2) Broader state coverage gives a Pareto improvement: expanding the initial-pose pool from 353 to 500 improves all metrics.

Q2. Common scored set, recovery denominators, and a shared knockdown subset

Thank you for pointing out the potential interpretation bias from differing evaluation denominators. To address the concern that different disturbance resistance leads to different knockdown subsets across models, we explicitly report the complete evaluation counts and add a comparison on a strictly matched shared knockdown subset (Table P).

(a) Common motion set and actual denominators

All models are evaluated on exactly the same benchmark. The common no-push scored set contains 118 motions, with 10 individually matched external pushes per motion, following Section 4.1 and Appendices B.2/B.4 with random seed 20260903. Push magnitude, direction, application point, and timing are matched across all models, giving 1180 disturbed trials. The actual denominator of each model's original conditional recovery rate—its number of actual knockdown trials—is reported in the second column of Table P.

(b) Recovery on exactly the same knockdown trials

For a fully fair comparison, we take the strict intersection of trials in which all eight models are knocked down, yielding 100 disturbed trials. We evaluate recovery using the original continuous one-second motion recovery criterion.

Table P. Original conditional recovery and recovery on the shared knockdown subset.

ModelOriginal conditional-recovery denominator (knockdowns)Original motion recovery (%) ↑Recovery on the shared knockdown subset
Pretrained MLP8090.000/100(0%)
Pretrained Transformer9050.000/100(0%)
Recovery exploration (all environments allocated to recovery)81953.7054/100(54%)
Split-environment exploration80061.8062/100(62%)
BFM-Zero49436.8037/100(37%)
STUBBORN (post-training reproduction)82442.1042/100(42%)
StableBFM(MLP)81473.6074/100(74%)
StableBFM(Transformer)95174.0073/100(73%)

On the identical shared knockdown subset, with denominator differences eliminated, StableBFM (MLP) still achieves 74% recovery, exceeding the next-best baseline, split-environment exploration (62%), and STUBBORN (42%) by 12 and 32 percentage points, respectively. This controlled comparison confirms that StableBFM's performance advantage comes from its recovery mechanism rather than evaluation-selection bias.

Q3. Teacher acquisition cost, amortization, and matched STUBBORN comparisons

Thank you for raising this point. Below, we analyze teacher-cost amortization and verify baseline fairness using strictly controlled ablation experiments.

(a) Teacher acquisition cost and reuse across students

StableBFM follows an efficient paradigm of training a teacher once and reusing it for multiple low-cost student upgrades through distillation. Even after fully accounting for teacher training—the BFM-Zero reproduction publicly provided through UFO costs approximately 4.5 GPU-hours, and MLP student post-training costs 2.0 GPU-hours—the total end-to-end cost of the first upgrade is only 6.5 GPU-hours, below the 7.50 GPU-hours required for a complete 4000-step STUBBORN post-training run. More importantly, the teacher policy is fully reusable in subsequent iterations, so each additional student upgrade costs only 2.0 GPU-hours, or 26.7% of STUBBORN's cost. Deployment retains only the original student decoder, with the same model capacity and control interface as the baseline and no additional inference overhead.

(b) Controlling training terrain and optimization steps

In the original experiments (Table C(a)), we match the pretrained tracker, motion pool, trainable modules, and evaluation protocol. We retain the original rough-terrain configuration and provide four times as many optimization steps for STUBBORN to explore recovery from the outset (4000 vs. 1000 steps), allowing sufficient exploration budget. To directly address whether terrain and optimization-step differences dominate the comparison, we add a controlled experiment that matches both flat-ground training and a 1000-step optimization budget (Table C(b)).

Table C(a). Original StableBFM and STUBBORN post-training configurations, following Appendix E.

SettingStableBFM(MLP)STUBBORN (post-training reproduction)
Robot and control interface29-DoF G1; original PD, observation, and tokenizer interfacesSame
Pretrained initializationOfficial SONIC MLP, step 41,550 checkpointSame
Training motion poolboneseed130kSame
Trainable parametersAction decoder (g1_dyn); all other modules frozenSame
Total parallel environments40964096
Checkpoint selectionFinal checkpoint at the end of the training budgetSame
Training terrainFlat groundRough terrain inherited from the pretrained tracker
Post-training optimization steps10004000

Table C(b). Controlled comparison with matched training terrain and post-training steps.

Method / configurationTraining terrainPost-training optimization stepsMotion Decode tracking success (%) ↑Eventual stand-up success (%) ↑Motion recovery (%) ↑
STUBBORN (original configuration)Rough terrain400073.9283.8542.10
STUBBORN (matched ablation)Flat ground100075.1882.2942.21
StableBFM (ours)Flat ground100076.8593.2373.60

Table C(b) clearly shows that when STUBBORN uses the same flat terrain and 1000-step optimization budget as StableBFM, its motion recovery remains at 42.21%, essentially unchanged from 42.10% with rough terrain and 4000 steps. A substantial gap of 31.39 percentage points remains relative to StableBFM's 73.60%. This demonstrates that StableBFM's effectiveness stems from its sample efficiency and guided recovery mechanism, rather than terrain or training-step bias.

Additional response to Weakness 2: real-robot reliability

To address real-robot deployment reliability, we provide the complete continuous recording corresponding to the hardware recovery sequence in Section 4.5: real-robot recovery video. The frame sequence in the paper is extracted from this original recording, which shows the full process of the G1 robot being disturbed, falling, autonomously adjusting its posture, and returning to a stable standing position.

To provide quantitative evidence of hardware robustness while accounting for joint wear and motor heating caused by repeated falls, we conducted 10 consecutive independent fall-recovery trials on the G1 under protective conditions. StableBFM stood up autonomously in 9 trials, achieving 90% success and a mean recovery time of 1.3 seconds, measured from torso contact with the ground until stable standing.

MethodTest conditionsTrialsAutonomous stand-up success (%)Recovery time (s)
StableBFMG1 hardware tests under protective conditions1090%(9/10)1.3

The continuous hardware video and repeated quantitative tests with a high success rate jointly verify StableBFM's transfer robustness and execution reliability in the physical world. We hope these additional experiments and results fully address your concerns.

Reviewer_oiAXOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

StableBFM demonstrates substantial recovery gains, while the evidence for preserving tracking, the necessity of its components, and reliable real-world performance remains incomplete. The method post-trains a SONIC decoder using tracking PPO alongside recovery imitation from a BFM-Zero expert queried on learner-visited states. It keeps the deployed architecture unchanged and introduces history perturbations for a Transformer variant. The core approach is technically plausible. The strongest result is the MLP: eventual stand-up success increases from 0% to 93.2%, and disturbance motion recovery reaches 73.6%, 1with two GPU-hours of post-training. Training in Isaac Lab and evaluating in MuJoCo strengthens the evidence. Distinguishing standing up from returning to an advancing reference is also valuable. These are meaningful results.

Critical strengths
  1. Practical problem and simple solution. It adds fall recovery to a pretrained humanoid controller without increasing deployed model capacity or requiring a separate recovery controller.
  2. Substantial recovery gains. The MLP achieves 93.2% eventual stand-up success and 73.6% motion recovery, compared with zero recovery for the pretrained tracker.
  3. Meaningful recovery objective. Evaluating return to a still-advancing motion reference goes beyond merely standing up.
  4. Encouraging deployment evidence. Isaac Lab training, MuJoCo evaluation, and a real G1 demonstration support transferability. Two GPU-hours of MLP post-training is attractive when an expert is available.
Critical weaknesses
  1. The experiments do not separate the benefit of expert supervision from the contribution of the proposed method. StableBFM receives supervision from an expert that already possesses recovery capability, whereas the main exploration baselines receive no expert supervision. Moreover, Appendix E.2 disables recovery PPO and instead uses reward-weighted self-imitation. These comparisons establish that expert-guided post-training is effective, but do not isolate the benefits of DAgger, the two-branch design, or asynchronous updates. A particularly informative comparison would use the same expert for offline BC + tracking PPO versus DAgger + tracking PPO. This also affects the novelty claim: state–reference discrepancy is used for interpretation and visualization, without directly influencing the objective or sampling procedure. Its conceptual value would be stronger if it led to a distinct, experimentally verified insight.

  2. The central claim of adding recovery while preserving tracking is not supported by a sufficient tradeoff analysis. On Motion Decode, MLP tracking success decreases from 82.38% to 76.85%, local MPJPE increases from 53.35 to 63.82 mm, and global MPJPE increases from 547 to 1182 mm (Table 2). Several metrics therefore show meaningful degradation. These costs may be acceptable given the recovery gains. However, the relevant question is whether the method achieves better recovery at a comparable level of tracking retention. A single configuration and an aggregate Stability Score cannot establish this. Varying the recovery-data fraction or imitation weight to report a tracking–recovery tradeoff curve would directly test the central claim.

  3. The recovery benchmark does not fully establish sustained motion resumption, and its fall definition needs validation. Appendix B.2 classifies a trial as a knockdown whenever root height falls below 0.55 m. For reference motions containing crouching or kneeling, successful tracking could potentially satisfy this condition. Filtering for successful unperturbed rollouts does not automatically exclude such motions. Likewise, requiring recovery above 0.65 m can conflict with a legitimate low-posture reference. This is a potential protocol issue, not evidence that the reported results are contaminated. The authors should characterize the selected motion set and verify whether unperturbed trajectories trigger the fall detector. Additionally, motion recovery requires only one successful one-second window. It does not require sustained tracking afterward. Consequently, 73.6% motion recovery should not be interpreted as 73.6% successful completion of disturbed tasks. Post-recovery tracking error, relapse rate, and eventual task completion would better support the claimed capability.

  4. The proposed explanation for the Transformer’s behavior is not causally established. The paper attributes recovery difficulties to conflicting historical context, but cache substitution could also help through ordinary regularization. The MLP and Transformer also use different pretrained implementations, preventing their performance difference from isolating the effect of history. In Table 3, dropping history improves stand-up success but reduces motion recovery, while substitution produces different effects. These results show that context interventions matter, but do not establish the proposed mechanism. Controlled interventions on history while holding the current state, goal, and model fixed would provide stronger evidence.

Questions for authors
  1. What is the specific benefit of DAgger over standard expert distillation? Can you compare against offline BC using the same recovery expert, tracking PPO branch, and training budget?

  2. How strong is the tracking–recovery tradeoff? Given the Motion Decode success decrease from 82.38% to 76.85%, can adjusting the recovery-data fraction or BC weight preserve more tracking capability while retaining substantial recovery gains?

  3. Does the recovery metric establish genuine, sustained motion resumption? How do you exclude intentional low postures from the height-based fall detector, and how often does successful tracking persist beyond the one-second recovery window?

5 comments

We sincerely thank you for your thoughtful observations, valuable insights, and the time and care you have devoted to reviewing our work. Below, we address each of your questions and concerns in turn.

Q1. Contributions of expert supervision, DAgger, and the training design

Thank you for your question. We agree that the key issue is not simply whether expert-guided post-training works, but whether online DAgger and separate PPO/BC updates provide additional performance gains when the same recovery expert is used. We therefore added direct comparisons of offline expert BC + tracking PPO, joint PPO–BC, and StableBFM, using the same teacher, tracking PPO branch, student initialization, and equivalent training budgets (Table A).

(a) The contribution of online DAgger

With the same teacher, offline BC achieves Motion Decode tracking success, eventual stand-up success, and motion recovery of 57.36% / 79.02% / 62.85%, substantially below StableBFM's 76.85% / 93.23% / 73.60%. This shows that the improvement comes not only from having expert supervision, but also from DAgger's continuous online correction of the states the student actually visits during recovery exploration, which better addresses distribution shift in recovery.

(b) The necessity of the asynchronous update design

Joint PPO–BC also performs below StableBFM, with corresponding rates of 58.78% / 72.14% / 65.53%. This shows that separate PPO and BC updates are not an optional implementation detail, but a training design necessary to obtain the current recovery gains while preserving tracking.

Table A. Ablations using the same recovery expert and budget.

MethodRecovery supervision dataUpdate schemeTracking success:AMASS / Motion Decode (%) ↑Clean / Eventual stand-up (%) ↑Motion recovery (%) ↑
Offline expert BC + tracking PPOFixed expert trajectoriesSame separate-update schedule as StableBFM60.91 / 57.3651.69 / 79.0262.85
Joint PPO–BCOnline aggregation of teacher supervision on learner-visited statesJoint-loss updates64.37 / 58.7846.44 / 72.1465.53
StableBFM(MLP)Online aggregation of teacher supervision on learner-visited statesSeparate PPO and BC updates, K=682.80 / 76.8560.42 / 93.2373.60

(c) The role of state–reference discrepancy

We use discrepancy as an analytical basis for explaining why this online training design is necessary, rather than as an additional optimization term. The pretrained tracker's state distribution is concentrated near the reference, whereas recovery must cover regions with much larger discrepancies. Offline expert trajectories alone are therefore insufficient for teaching the student to return from these states to the tracking region; supervision must continue on recovery states that the student actually visits. Figure 4 illustrates this relationship: the additional recovery data cover large-discrepancy regions, and the post-trained model can return from those regions to the tracking region. We will make this connection more explicit in the revised manuscript to avoid presenting discrepancy as a separate optimization module.

Q2. Tracking–recovery tradeoffs and training configurations

Following your suggestion, we compared the effects of the recovery-environment fraction, BC weight, and initial-pose pool size. The default configuration achieves Motion Decode tracking success / eventual stand-up / motion recovery of 76.85% / 93.23% / 73.60%.

(a) Interpreting the increase in global MPJPE

To address the increase in global MPJPE, we conducted a controlled data-level sensitivity analysis. Starting from ideally tracked trajectories, we injected identical global errors into the 1503 Motion Decode clips and a stratified set of 1500 AMASS clips reconstructed according to the appendix protocol (seed 20260902). Table D summarizes the results.

Table D. Dataset characteristics and controlled sensitivity to global errors.

Data characteristic / injected errorReported metric (unit)Motion Decode (1503 clips)AMASS (1500 clips)
Clip durationMedian (s)23.16.7
Maximum displacement from the starting pointMedian (m)0.410.71
Displacement from the starting point exceeding 3 mFraction of clips (%)19.25.1
Displacement from the starting point exceeding 5 mFraction of clips (%)13.30.9
Displacement from the starting point exceeding 10 mFraction of clips (%)6.10.0
Cumulative turningMedian (°)305125
Heading drift of 1°/sstrict-foot SR(%)5581
Random walk, σ=0.05 m/√sstrict-foot SR(%)1249
Fixed heading offset of 3°strict-foot SR(%)8498
Stride-length error of 5%strict-foot SR(%)8498
Fixed heading offset of 10°strict-foot SR(%)6855
Heading drift of 1°/sMPJPE-G increase (mm)+293+46

SR is the strict-foot diagnostic success rate after data-level error injection, using a 0.20 m threshold on the world-coordinate distance between corresponding feet. Tracking success in Table 2 follows the original no-foot evaluation. MPJPE-G increases are measured relative to the ideal trajectories before error injection.

Table D shows that Motion Decode has longer clips and a heavier tail of long-distance motions. Under the same heading drift of 1°/s, random walk, fixed heading offset of 3°, and stride-length error of 5%, its strict-foot SR is lower than AMASS. Most relevantly, heading drift of 1°/s increases MPJPE-G by 293 / 46 mm on MD / AMASS, so the same global-error process produces a larger metric increase on MD. We therefore assess tracking costs through both global trajectory error and aligned pose error: in Table 2, MLP global MPJPE increases by 634.60 mm, while local and PA errors increase by 10.47 and 5.43 mm. The configuration comparison below reports the effects on tracking and recovery.

(b) Recovery-environment fraction, BC weight, and initial-state coverage

The additional experiments show that weakening recovery supervision generally reduces recovery performance. Reducing recovery environments from 1024 to 512 gives 76.21% tracking success, while eventual stand-up and motion recovery decline to 91.48% and 71.12%. Reducing the BC coefficient from 2 to 1 gives 76.58% tracking success, with eventual stand-up / motion recovery falling to 91.72% / 70.94%.

Increasing recovery supervision does not yield a more pronounced overall improvement: with 2048 recovery environments, the three rates are 76.69% / 93.52% / 73.92%, close to default. Raising the BC coefficient to 4 reduces tracking success to 75.96%, without improving recovery (93.47% / 73.18%). Expanding the initial-pose pool from 353 to 500 yields 76.91% / 93.35% / 73.79%, a small improvement over default.

These comparisons clarify the tradeoff. The default configuration is not the only viable choice, but remains a balanced operating point within the tested range. Expanding the initial-pose pool yields a modest improvement, suggesting room for further optimization. Among the tested configurations, however, we did not observe an alternative that substantially preserves more tracking while maintaining comparable recovery gains.

Hyperparameter ablations (single-factor variations).

ConfigurationRecovery environments / total environmentsInitial posesBC coefficientTracking success (%)Local / global MPJPE (mm)Eventual stand-up success (%)Motion recovery (%)
Default1024 / 4096353276.8563.82 / 1181.9093.2373.60
Fewer recovery environments512 / 4096353276.2164.35 / 1194.7291.4871.12
More recovery environments2048 / 4096353276.6963.98 / 1186.0293.5273.92
Expanded initial-pose pool1024 / 4096500276.9163.74 / 1180.2793.3573.79
Lower BC coefficient1024 / 4096353176.5864.08 / 1188.4191.7270.94
Higher BC coefficient1024 / 4096353475.9665.21 / 1206.8593.4773.18

Q3. Low-posture classification and sustained tracking after recovery

Thank you for identifying these two important details of the evaluation protocol.

(a) What the motion recovery criterion establishes

Our motion recovery metric directly tests the robot's ability to reconnect with a dynamic reference trajectory after falling. Following a disturbance, the reference motion continues to advance at its original pace. Successful recovery requires the robot to remain stably upright for 50 consecutive control steps (1 second), while its body height, global orientation, end-effector heights, and whole-body local joint pose satisfy tight reference-relative tracking thresholds. The 73.60% motion recovery rate therefore measures actual closed-loop resumption of dynamic reference tracking.

(b) Verification of intentional low postures

We checked the reference trajectories and unperturbed execution records of all motions in the common scored set. Within the same observation interval used for the disturbance evaluation, 0 motions triggered the 0.55 m knockdown threshold during normal execution, and 0 motions had low-posture references that conflicted with the 0.65 m recovery-height condition. The check shows that none of the unperturbed trajectories in the current scored motion set triggered the 0.55 m knockdown threshold, and legitimate low-posture references did not conflict with the recovery-height requirement. The original recovery statistics were therefore not systematically affected by low-posture motions. This directly verifies that normal crouching or kneeling motions did not introduce systematic bias into the original scoring.

(c) Sustained execution beyond the one-second recovery window

After a trial first satisfied the 1-second recovery criterion, we continued observing the same trajectory to measure subsequent tracking performance, as reported below. Among all actual knockdown trials, 69.45% maintained successful tracking for at least 3 additional seconds, and 66.12% maintained successful tracking through the end of the episode. Conditioned on having recovered, these correspond to 94.36% continuing successfully for at least 3 seconds and 89.84% continuing successfully through episode end, with a subsequent tracking failure rate of 10.16%.

As a reference for overall tracking capability, the same post-trained MLP achieves 82.80% nominal tracking success and 17.20% failure on the full AMASS evaluation set in the paper; the above 89.84% measures continued execution over the remaining portions of recovered trials. Together, these results support that the robot returns to stable, normal tracking after recovery: the vast majority of recovered trials continue successfully to the end, demonstrating effective integration of recovery skills and the original tracking capability within the same decoder.

Evaluation criterion (MLP)**Success rate (%) **
Original criterion: recovery conditions satisfied for 1 consecutive second73.60
Successful tracking for at least 3 additional seconds after initial recovery69.45
Successful tracking through the end of the episode after initial recovery66.12

All three rates in the table use all actual knockdown trials as the denominator.

Reviewer_iE5pOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents StableBFM, a post-training framework that extends pretrained humanoid behavior foundation models (BFMs) with fall-recovery capabilities. The framework combines the original PPO-based tracking objective with DAgger-style behavior cloning on learner-visited recovery states, in the asychronous update manner, supervised by a pretrained BFM-Zero expert. Experiments on the Unitree G1 demonstrate substantial improvements in simulated fall recovery.

Critical strengths
  • Enabling pretrained humanoid motion-tracking policies to recover from severe disturbances without switching to a separate recovery controller is an important challenge. The motivation is clear, and the proposed fixed-capacity adaptation setting is practically relevant.
  • The post-trained BFM demonstrated strong failure recovery, while preserving the nominal tracking performance.
Critical weaknesses
  • The paper lacks comparisons against simpler alternatives, particularly tracking PPO combined with offline expert BC, and the standard joint PPO-BC joint optimization reward objective. I would appreciate if the authors can provide these ablations to demonstrate the effectiveness of the asychronous update of PPO and BC.
  • Missing qualitative visual videos of real-world G1 deployment, and it is hard to evaluate the tracking quality and recover ability in real-world. Also, the paper doesn't include quantitative evaluation of real world recovery, like the recovery success rate and recovery time evaluated in related paper [1].

[1]. Ren, Xiao, et al. "Stubborn: A Streamlined and Unified Reinforcement Learning Framework for Robust Motion Tracking and Fall Recovery for Humanoids." arXiv preprint arXiv:2606.12814 (2026).

Questions for authors

Listed in the weaknesses

4 comments

We sincerely thank you for your thoughtful observations, valuable insights, and the time and care you have devoted to reviewing our work. Below, we address each of your questions and concerns in turn.

Q1. Comparisons with offline expert BC and joint PPO–BC

Thank you for your comments. Following your suggestion, we added ablation comparisons using the same teacher, student initialization, and equivalent training budgets. The results are reported in Table A.

Table A. Comparisons with the same teacher, student initialization, and training budget.

MethodRecovery supervision dataUpdate schemeTracking success:AMASS / Motion Decode (%) ↑Clean / Eventual stand-up (%) ↑Motion recovery (%) ↑
Offline expert BC + tracking PPOFixed expert trajectoriesSame separate-update schedule as StableBFM60.91 / 57.3651.69 / 79.0262.85
Joint PPO–BCOnline aggregation of teacher supervision on learner-visited statesJoint-loss updates64.37 / 58.7846.44 / 72.1465.53
StableBFM(MLP)Online aggregation of teacher supervision on learner-visited statesSeparate PPO and BC updates, K=682.80 / 76.8560.42 / 93.2373.60

(a) Offline BC versus online DAgger

With the same teacher, offline BC achieves eventual stand-up and motion recovery rates of 79.02% / 62.85%, whereas DAgger reaches 93.23% / 73.60%. This shows that continuously correcting the states the student encounters during actual execution helps recovery skills transfer more effectively to closed-loop execution.

(b) Joint versus separate PPO–BC updates

With online teacher supervision fixed, joint PPO–BC achieves tracking, eventual stand-up, and motion recovery rates of 58.78% / 72.14% / 65.53%, compared with 76.85% / 93.23% / 73.60% for separate updates (tracking is evaluated on Motion Decode). This comparison shows that separate PPO and BC updates help preserve existing tracking performance while achieving better recovery gains. The role of the independent update schedule is also evident in the K ablation: increasing K from 1 to 6 raises motion recovery from 54.90% to 73.60%, while AMASS tracking success remains at 82.67%–82.80%.

Q2. Real-robot video, recovery success, and recovery time

(a) Continuous G1 recovery video

We provide the video corresponding to the real-robot stand-up sequence in Section 4.5: real-robot recovery video. The frame sequence in the paper is taken from this recording, which continuously shows the robot falling after a disturbance and standing up autonomously.

(b) Quantitative hardware evaluation

To provide quantitative evaluation, we conducted 10 G1 recovery trials under protective conditions to limit joint wear and motor heating from repeated falls. StableBFM stood up autonomously in 9 trials, achieving a 90% success rate and a mean recovery time of 1.3 seconds, as reported below.

MethodTest conditionsTrialsAutonomous stand-up success (%)Recovery time (s)
StableBFMG1 hardware tests under protective conditions109/10(90%)1.3

Recovery time is measured from the robot's torso contacting the ground until it regains a stable standing posture. The video and quantitative tests jointly demonstrate the model's recovery capability and execution performance in the real world.

Dear Reviewer iE5p, Thank you for your careful review. We summarize the additional results below.

  1. Online DAgger. With the same teacher and budget, online DAgger improved motion recovery from 62.85% to 73.60% over offline expert BC.
  2. Separate PPO and BC updates. With the same online supervision, separate updates improved both tracking and recovery over a joint loss.
  3. G1 hardware. The supplementary video shows autonomous stand-up, which succeeded in nine of ten trials under protective conditions. We shared the same evidence with Reviewer S6rF, who noted that our rebuttal addressed their concerns.

We would be happy to clarify any remaining points before the discussion closes. Thank you again for your time and consideration.

Thanks authors for the detailed rebuttal. My concerns have been addressed, and I will raise the score.

Thank you for taking the time to revisit our paper. We’re glad our response addressed your concerns.

Reviewer_xfCROpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes StableBFM, a post-training framework that adds fall recovery to a pretrained motion-tracking controller without increasing its deployed model capacity. Motivated by the limited coverage of large state–reference discrepancies during tracking pretraining, the method combines continued tracking PPO updates with DAgger-style recovery supervision from a BFM-Zero teacher. For a context-conditioned decoder, the authors additionally introduce KV-cache reconstruction and perturbation. Experiments evaluate nominal tracking, fall-to-stand recovery, disturbance recovery, and deployment on a Unitree G1.

Critical strengths
  1. The motivation is practically relevant. High-fidelity tracking and recovery from off-nominal states are both necessary for deployment, and extending an existing controller offers a useful alternative to training a separate recovery policy.
  2. The cache-management component is an interesting aspect of the method. Distinguishing numerical cache staleness from harmful dependence on historical context identifies a concrete issue in post-training context-conditioned controllers. The ablations provide evidence that cache perturbations affect recovery behavior.
  3. The method substantially improves recovery over the pretrained tracker while retaining a single deployed decoder. Evaluating return to an advancing reference, in addition to standing up, is valuable. The real-robot demonstration also supports the feasibility of the approach.
Critical weaknesses
  1. The conceptual contribution needs clearer positioning relative to motion in-betweening. Bridging the gap between the current state and a target state is closely related to motion in-betweening. More directly, HERACLES: Bridging Precise Tracking and Generative Synthesis for General Humanoid Control addresses the same tension between precise tracking and recovery under large state deviations. It uses a state-conditioned flow-matching middleware to generate transition references through closed-loop replanning.
  2. The generality and efficiency of the framework are conditional on access to a capable recovery teacher. The formulation permits a generic expert, but the empirical evidence uses only an reproduction of BFM-Zero. This demonstrates transfer from one particular teacher rather than establishing that the framework can broadly extend pretrained controllers. The important question is whether the method requires a large generalist teacher, or whether a smaller task-specific recovery policy could provide comparable supervision. The reported post-training cost is useful, but its practical interpretation depends on whether the teacher is already available. The authors should report the teacher-training cost separately and discuss when that cost can reasonably be amortized.
  3. The evidence does not consistently support the claim of preserving tracking performance. Table 2(a) shows meaningful tracking degradation relative to the corresponding pretrained models. For the Markov decoder, Motion Decode success decreases from 82.38% to 76.85%, although AMASS success remains nearly unchanged (83.01% to 82.80%). For the context decoder, success decreases from 79.51% to 71.32% on Motion Decode and from 81.73% to 78.93% on AMASS. Tracking errors also worsen: for example, the Markov decoder’s Motion Decode global MPJPE increases from 547.30 mm to 1181.90 mm. Recovery transfer is also incomplete relative to the teacher. BFM-Zero achieves 98.96% eventual and 79.17% clean fall-to-stand success, compared with 93.23% / 60.42% for the Markov student and 90.63% / 54.17% for the context student.
Questions for authors
  1. How does StableBFM compare with HERACLES in contribution and practical tradeoffs, given its fixed reference stream and unchanged deployed capacity?
  2. Can BFM-Zero be replaced by a smaller recovery expert, and how sensitive are the results to teacher quality?
  3. What level of tracking degradation is considered acceptable when claiming capability preservation?
  4. What evidence distinguishes the proposed history-shortcut explanation from a more general regularization effect of cache perturbation?
4 comments

We sincerely thank you for your thoughtful observations, valuable insights, and the time and care you have devoted to reviewing our work. Below, we address each of your questions and concerns in turn.

Q1. Contribution and practical tradeoffs relative to HERACLES

Thank you for your question. The two approaches address recovery at different levels: HERACLES dynamically generates transition references through a higher-level module for a low-level tracker to execute, whereas StableBFM extends the capabilities of the low-level tracker itself, enabling it to recover directly from large deviations and reconnect with the original reference.

These choices lead to different solutions. HERACLES adapts the reference at a higher level to accommodate the current state; StableBFM equips the low-level controller to handle large state deviations. To evaluate our contribution, we keep the reference stream and deployed capacity fixed and compare the same tracker before and after post-training, directly measuring the recovery gains introduced by StableBFM. In deployment, HERACLES relies on the higher-level module to organize recovery transitions, while StableBFM allows the existing system to continue using its original interface to invoke a unified controller with both tracking and recovery capabilities.

Q2. Smaller recovery experts, teacher quality, and training cost

(a) Replacing the recovery teacher

Thank you for your comments. We added an ablation that replaces the recovery teacher: keeping the pretrained student, post-training settings, and evaluation protocol unchanged, we replace BFM-Zero with the smaller task-specific recovery expert SuperMimic(recovery). The results show that the smaller teacher also provides effective supervision: the post-trained student achieves 70.33% motion recovery, compared with 73.60% using the BFM-Zero teacher. This demonstrates that StableBFM does not depend on a large generalist teacher; a smaller recovery-specific teacher is also effective.

Table T. Student recovery results with different teachers.

Recovery teacherTeacher parameters (M) ↓Student clean stand-up (%) ↑Student eventual stand-up (%) ↑Student motion recovery (%) ↑
BFM-Zero31.9M60.4293.2373.60
SuperMimic(recovery)8.89M55.9181.8070.33

(b) Teacher acquisition cost and amortization

We distinguish teacher acquisition cost from student post-training cost. Training the BFM-Zero reproduction used in this study costs approximately 4.5 GPU-hours, using Roboparty's UFO framework, while incremental student post-training costs 2.0 GPU-hours. Treating the teacher as a reusable pretrained asset, each subsequent iteration incurs a marginal cost of only 2.0 GPU-hours; including the initial upfront cost gives a total of 6.5 GPU-hours. Only the student decoder is retained at deployment, introducing no additional inference overhead. We will further clarify these two cost components in the revised manuscript.

Q3. Preservation of tracking capability after recovery injection

(a) Tracking costs and recovery gains

We consider the current tracking cost to be within a reasonable range: the model continues to complete most of the original tracking tasks without a marked deterioration in local motion quality, while acquiring substantial recovery capabilities. Specifically, AMASS tracking success remains at 82.80% (83.01% before post-training), and Motion Decode success is 76.85% (82.38% before post-training), a decrease of 5.53 percentage points. Meanwhile, local MPJPE increases by only 10.47 mm (53.35 → 63.82 mm), and PA error increases by only 5.43 mm (32.82 → 38.25 mm), while eventual stand-up success rises from 0 to 93.23% and motion recovery reaches 73.60%.

(b) Global trajectory error on Motion Decode

Post-training changes in global tracking accuracy incur different metric costs on the two datasets. For the same MLP, MPJPE-G increases from 606.50 to 682.60 mm on AMASS, an increase of 76.10 mm, and from 547.30 to 1181.90 mm on Motion Decode, an increase of 634.60 mm. Meanwhile, AMASS local error remains essentially unchanged (52.64 → 52.43 mm), MD local error increases by 10.47 mm, and their PA errors increase by 1.84 mm and 5.43 mm, respectively. The question is why the global-error increase on MD is much larger than the increase in aligned pose error.

There is a concrete trajectory-level mechanism for this amplification: persistent small velocity or heading deviations can accumulate into position drift over time, while heading-induced world-coordinate position errors also depend on the displacement covered by the motion. MD has longer clips than AMASS (median 23.1 s vs. 6.7 s), and more clips exceed 3 m displacement from their starting point (19.2% vs. 5.1%). To test this distribution-dependent amplification, we inject identical heading drift of 1°/s into ideal reference trajectories. MPJPE-G increases by 293 mm on MD and 46 mm on AMASS (Table D). This controlled result shows that the same global-error process produces a larger metric cost on MD, providing an explanation for the larger global-error increase observed there after post-training. Together with the small changes in local and PA errors, these results support that the model retains good motion execution quality, while changes in global tracking accuracy are more pronounced in the global metric on MD.

(c) Comparison with other recovery methods

Comparisons across methods show that StableBFM achieves better stand-up and recovery performance while retaining high tracking success. For example, relative to the split-environment baseline, which has a similar tracking success rate, eventual stand-up success improves from 89.58% to 93.23%, and motion recovery from 61.80% to 73.60%. StableBFM therefore achieves a more favorable tracking–recovery tradeoff.

Table D. Reported global-error changes and data-level sensitivity analysis.

ComparisonMotion Decode (1503 clips)AMASS (1500 clips)
Reported MPJPE-G before post-training (mm)547.30606.50
Reported MPJPE-G after post-training (mm)1181.90682.60
Reported increase in MPJPE-G after post-training (mm)+634.60+76.10
Median clip duration (s)23.16.7
Clips exceeding 3 m displacement from the starting point (%)19.25.1
Controlled data-level analysis: MPJPE-G increase under identical 1°/s heading drift (mm)+293+46

Q4. History shortcuts versus general cache regularization

To distinguish history substitution (Subst) from general regularization effects, we added a random cache-noise control (Table H). All three runs use the same initialization, teacher, and budget, changing only cache treatment during recovery BC. Random noise raises motion recovery from 68.50% with Default to 70.12%, while Subst further improves it to 74.00%, achieving better recovery at a similar tracking level.

These results show that random perturbations provide some regularization benefits. Subst, by using histories from other rollouts in the same batch, reduces dependence on a particular erroneous history while retaining temporal information, thereby supporting more effective motion resumption after recovery. This is also consistent with our proposed explanation of the history-processing mechanism.

Table H. Tracking and recovery under different cache treatments.

Cache treatment during recovery BCAMASS tracking success (%)Motion Decode tracking success (%)Clean stand-up (%)Eventual stand-up (%)Motion recovery (%)
Default79.7369.8435.9089.1068.50
Random cache noise79.1870.4146.7088.7670.12
Subst78.9371.3254.2090.6074.00

The authors have added recovery experiments in their rebuttal, which I believe address my concerns. I have therefore decided to raise my score.

Thank you for taking the time to read our rebuttal and follow up. Appreciated

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the StableBFM rebuttal

We thank the reviewers for their feedback. Below we summarize the main points from our replies and the evidence we added during the discussion.

  • Recovery within the tracker (xfCR). HERACLES generates transition references upstream. StableBFM post-trains the existing tracker to recover and rejoin the original, continuously advancing reference without changing the deployed model capacity or control interface. A smaller, recovery-specific teacher also proved effective, and only the student is used at inference.
  • What the training design contributes (iE5p, oiAX). With the teacher, initialization, and training budget held fixed, online DAgger outperformed offline expert BC under the same update schedule. With online supervision held fixed, separate PPO and BC updates outperformed a joint loss on both tracking and motion recovery.
  • Tracking cost and global drift (xfCR, oiAX, S6rF). AMASS tracking success was nearly unchanged (83.01% to 82.80%), while Motion Decode dropped from 82.38% to 76.85%. Applying the same heading drift to both datasets increased global position error by 293 mm on Motion Decode versus 46 mm on AMASS, showing that Motion Decode is more sensitive to accumulated drift.
  • Recovery criterion and matched comparisons (oiAX, S6rF). Recovery requires the robot to resume tracking the still-advancing reference after a fall. Normal low postures did not trigger the fall threshold on the scored motions. On the same 100 trials in which all eight models fell, StableBFM (MLP) recovered in 74, compared with 62 for split-environment exploration and 42 for STUBBORN. The gap over STUBBORN also persisted with matched training terrain and update counts.
  • History substitution versus random noise (xfCR, oiAX). Starting from the same Transformer checkpoint and changing only the cache treatment during recovery BC, history substitution achieved 74.00% motion recovery versus 70.12% with random cache noise, at similar tracking success.
  • G1 hardware evidence (iE5p, S6rF). We added a supplementary video (https://recovery-clip-a7e39d.genialswan15.chatgpt.site) showing the G1 robot falling after a disturbance and standing up autonomously. In ten hardware trials under protective conditions, the robot stood up autonomously nine times, with a mean recovery time of 1.3 seconds from torso contact to stable standing.

As the discussion period draws to a close, we’d be happy to answer any remaining questions before it ends. If there are no further questions, we’d appreciate hearing your final assessment in light of the discussion. Thank you again for your time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.