acceptodds
Desk rejected at ICLR 2027

Heracles: Bridging Precise Tracking and Generative Synthesis for General Humanoid Control

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Achieving general-purpose humanoid control requires a delicate balance between the precise execution of commanded motions and the flexible, anthropomorphic realignment of state and intent. Current general controllers predominantly formulate motion control as a rigid reference-tracking problem. This formulation implicitly assumes that the reference is always reachable from the current state; once this assumption breaks, these trackers produce myopic, single-step corrections and exhibit brittle, non-anthropomorphic failure modes, lacking the generative adaptability inherent to human motor control. To overcome this limitation, we propose Heracles, a novel state-conditioned generative middleware that bridges precise motion tracking and generative synthesis. Rather than relying on rigid tracking paradigms or explicit mode-switching between a tracker and a dedicated recovery policy, Heracles operates as an intermediary layer between high-level reference motions and low-level physics trackers. By conditioning on the robot's real-time state, the flow-matching model implicitly adapts its behavior: it approximates an identity map when the state closely aligns with the reference, preserving zero-shot tracking fidelity. Conversely, when encountering significant state deviations, it seamlessly transitions into a generative synthesizer, producing an anthropomorphic whole-body maneuver back toward the reference. Our framework demonstrates that integrating generative priors into the control loop not only significantly enhances robustness against extreme perturbations but also elevates humanoid control from a rigid tracking paradigm to a generative general-purpose architecture. Videos and demos are available at https://heracles-humanoid-control-submission.github.io

Primary area:
applications to robotics, autonomy, planning
Keywords:
humanoid robots, whole-body control, motion tracking

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_Qp3BOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

Heracles addresses humanoid tracking when the reference differs substantially from the current state. It places a state-conditioned flow-matching model between the reference and a low-level RL tracker. Given the robot’s current configuration and one reference frame, the model generates a 0.2s residual trajectory with eight keyframes and replans at 25Hz. The tracker uses an iFSQ representation and is trained with PPO and adaptive motion sampling. Experiments include tracking and recovery tests in MuJoCo, transfer to SONIC, comparison with HumanUP, and demonstrations on a Unitree G1.

Critical strengths
  1. The middleware only changes the reference input and can be attached to existing trackers without retraining. It improves SONIC’s recovery results from 0/40 to 18/40 and runs in real time onboard.
  2. With the same tracker, adding the middleware raises Discrete CR from 26.3% to 92.0%. Recovery using a single standing reference improves from 0/40 to 37/40 configurations.
Critical weaknesses
  1. The middleware is a conditional motion in-betweening model trained offline on motion capture and separately from the tracker. It connects to the tracker only through reference inputs. On Recovery and Discrete, all baselines directly track the raw reference, without an interpolation baseline.
  2. iFSQ does not stand out among the baselines. Under identical observations, VQ-VAE achieves higher CR (91.9% vs 86.0%, and 67.1% vs 39.4% on Discrete), although the paper aims to improve success in difficult cases.
  3. The 0.2s planning horizon and eight-keyframe window are not studied experimentally. Adaptive motion sampling is used only for tracker training and is listed as a contribution, but no ablation is provided.
  4. The paper does not specify the source sequences of the Discrete keyframes or how long each keyframe is held before switching. It also does not explain how the eight planned frames over 0.2s are passed to the 50Hz tracking policy.
Questions for authors

No other questions.

3 comments

(1/3)

We thank the reviewer for the detailed reading of the method and the experiments. Each weakness raised in the review is addressed below, together with the additional experiments we ran for it. Our response is split into three comments.

W1. Interpolation baseline on the recovery and discretized subsets

The middleware and the tracker are trained separately and communicate only through the reference. This separation is what allows the same middleware to drive both our tracker and SONIC without retraining. To measure how much the learned model contributes beyond supplying intermediate poses, we built an interpolation planner that takes the place of the flow matching network and leaves every other component of Heracles in place.

At each replanning cycle, the interpolation planner places the eight keyframes on a straight path from the current state to the target. Joint angles and root position change linearly along this path, and root orientation changes by spherical linear interpolation. The first keyframe equals the current state, and the root horizontal position is expressed relative to the current root, both as in Heracles. The state representation, the 0.2 s horizon with K = 8, replanning at 25 Hz, the target given to the planner, the conversion of keyframes into the 50 Hz reference, the tracker checkpoint, the evaluation code and all metric definitions are identical to Heracles.

Table R1. Recovery and discretized subsets of Table 1, same iFSQ tracker, CR in percentSubsetPlannerCR↑Joint (rad)↓Height (m)↓Ori (rad)↓LinVel (m/s)↓RecoveryNone92.21.4410.0680.3230.175Interpolation91.11.2680.0680.2810.151Flow matching, Heracles97.81.1000.0250.1430.173DiscreteNone26.32.2080.4511.3440.258Interpolation15.12.7780.5181.4530.201Flow matching, Heracles92.01.1810.0550.1590.314\begin{array}{c}\textbf{Table R1. }\text{Recovery and discretized subsets of Table 1, same iFSQ tracker, CR in percent}\cr\begin{array}{|l|l|c|c|c|c|c|}\hline\text{Subset}&\text{Planner}&\text{CR}\uparrow&\text{Joint (rad)}\downarrow&\text{Height (m)}\downarrow&\text{Ori (rad)}\downarrow&\text{LinVel (m/s)}\downarrow\cr\hline\text{Recovery}&\text{None}&92.2&1.441&0.068&0.323&0.175\cr&\text{Interpolation}&91.1&1.268&0.068&0.281&\color{red}{0.151}\cr&\text{Flow matching, Heracles}&\color{red}{97.8}&\color{red}{1.100}&\color{red}{0.025}&\color{red}{0.143}&0.173\cr\hline\text{Discrete}&\text{None}&26.3&2.208&0.451&1.344&0.258\cr&\text{Interpolation}&15.1&2.778&0.518&1.453&0.201\cr&\text{Flow matching, Heracles}&\color{red}{92.0}&\color{red}{1.181}&\color{red}{0.055}&\color{red}{0.159}&0.314\cr\hline\end{array}\end{array}

Table R2. Autonomous recovery from the 40 fallen configurations of Table 10, same iFSQ trackerPlannerSR↑None0/40Interpolation0/40Flow matching, Heracles37/40\begin{array}{c}\textbf{Table R2. }\text{Autonomous recovery from the 40 fallen configurations of Table 10, same iFSQ tracker}\cr\begin{array}{|l|c|}\hline\text{Planner}&\text{SR}\uparrow\cr\hline\text{None}&0/40\cr\text{Interpolation}&0/40\cr\text{Flow matching, Heracles}&\color{red}{37/40}\cr\hline\end{array}\end{array}

On the recovery subset the reference is continuous, and the robot mostly stays close to it. There, interpolation performs similarly to the tracker alone, with 91.1% CR against 92.2% and somewhat lower joint and orientation errors. Heracles reaches 97.8% CR with the lowest joint, height and orientation errors. The difference widens when the gap between the robot and the reference is large. On the discretized subset, interpolation lowers CR from 26.3% to 15.1%, below the tracker without a planner, and it recovers none of the 40 fallen configurations, where Heracles recovers 37. Two features of straight-line planning account for this.

  • Each interpolated plan reaches the target within 0.2 s, whatever the distance. After a switch in a discretized reference, the pelvis has to move by 0.37 to 0.90 m, as reported in Sec. 4.1. The interpolation planner asks the tracker to complete this motion within a single 0.2 s horizon. The flow matching middleware predicts the motion increment for the next 0.2 s, as described in Sec. 3.2, and this increment stays bounded regardless of how far away the target is.
  • Getting up passes through support postures that lie off the straight path. Humans rise from the ground by rolling onto the side, pushing up on the arms and kneeling. These intermediate postures are not on the straight path between lying and standing, and some of them move temporarily away from the standing target. The middleware learns these postures from human motion capture and reproduces them in its plans, while a straight-line plan skips them.

The updated supplementary material contains videos of the interpolation planner on 10 fallen configurations randomly chosen from Table 10, and this comparison is reported in App. I of the revised paper.

(2/3)

W2. Reference fidelity of the iFSQ tracker and the higher CR of VQ-VAE

Within Heracles, the middleware handles large deviations from the reference, and the tracker executes the planned reference. We therefore chose the tracker by how closely it follows a reference. CR measures this only partly, since it counts a frame as complete whenever root height stays within 0.3 m and root orientation within 1.2 rad of the reference, so a policy that keeps its balance while following the joints loosely still reaches a high CR. Joint error measures this directly.

Table R3. CR in percent and joint error in rad from Table 1SubsetMetricVQ-VAEiFSQ-BMiFSQHeraclesOverallCR↑91.986.085.294.0Joint↓1.8831.3391.1090.969DailyCR↑97.496.296.393.7Joint↓1.6740.9450.8020.895RecoveryCR↑90.884.892.297.8Joint↓2.4241.8551.4411.100DiscreteCR↑67.139.426.392.0Joint↓2.3202.6612.2081.181\begin{array}{c}\textbf{Table R3. }\text{CR in percent and joint error in rad from Table 1}\cr\begin{array}{|l|l|c|c|c|c|}\hline\text{Subset}&\text{Metric}&\text{VQ-VAE}&\text{iFSQ-BM}&\text{iFSQ}&\text{Heracles}\cr\hline\text{Overall}&\text{CR}\uparrow&91.9&86.0&85.2&\color{red}{94.0}\cr&\text{Joint}\downarrow&1.883&1.339&1.109&\color{red}{0.969}\cr\hline\text{Daily}&\text{CR}\uparrow&97.4&96.2&96.3&93.7\cr&\text{Joint}\downarrow&\color{blue}{1.674}&0.945&{0.802}&0.895\cr\hline\text{Recovery}&\text{CR}\uparrow&90.8&84.8&92.2&\color{red}{97.8}\cr&\text{Joint}\downarrow&2.424&1.855&1.441&\color{red}{1.100}\cr\hline\text{Discrete}&\text{CR}\uparrow&67.1&39.4&26.3&\color{red}{92.0}\cr&\text{Joint}\downarrow&\color{blue}{2.320}&2.661&2.208&\color{red}{1.181}\cr\hline\end{array}\end{array}

VQ-VAE has the highest overall joint error of all trackers in Table 1, 1.883 rad. Its comparison with iFSQ-BM isolates the quantizer, since the two share the same encoder-decoder policy and observations and differ only in the latent. iFSQ-BM reduces overall joint error by 29% relative to VQ-VAE, and iFSQ with our observations by 41%. The gap is present even on smooth daily motions, where the joint error of VQ-VAE, 1.674 rad, is about 1.8 times that of iFSQ-BM and 2.1 times that of iFSQ.

We attribute both the higher CR and the higher joint error of VQ-VAE to how it quantizes the reference. VQ-VAE replaces each reference window with the nearest vector of a learned codebook, so the tracker executes a nearby prototype motion rather than the reference itself. iFSQ instead quantizes each latent channel independently on a fixed grid, so the number of representable codes grows exponentially with the latent dimension and the reference is represented at a much finer resolution. Rounding to prototypes limits precision, which explains the gap on daily motions. It also makes the tracker conservative, because abrupt or unusual parts of the reference are smoothed toward familiar motions and the robot keeps its balance more often. This effect is largest on the discretized subset, where VQ-VAE reaches 67.1% CR, while its joint error of 2.320 rad remains above the 2.208 rad of iFSQ and nearly twice the 1.181 rad of Heracles.

In Heracles, smoothing an infeasible reference into a feasible transition is the task of the middleware, and the tracker is expected to execute the planned transition as given. A tracker that rounds its input to a fixed set of prototypes would also distort the planned transition. We therefore use iFSQ, which has lower joint error than VAE, VQ-VAE, SONIC, iFSQ-BM and iFSQ+H in every subset and the lowest joint error of all standalone trackers on the discretized subset.

W4. Construction of the discretized references and the planner-tracker interface

Discretized references. All 150 discretized references come from fall-and-get-up recordings in the proprietary motion-capture collection described in Sec. 4.1, with no overlap with the daily or recovery subsets. Each reference consists of three static poses from the same recording, each a single stable frame with minimal joint motion, and consecutive poses differ in pelvis height by at least 0.30 m. Every reference lasts 550 frames at 50 fps, i.e. 11 s, and each pose stays fixed over its segment.

  • Frames 0-49 hold the standing pose, 50 frames or 1 s.
  • Frames 50-149 hold the fallen pose, 100 frames or 2 s.
  • Frames 150-549 hold the risen pose, 400 frames or 8 s.

The reference jumps from standing at frame 49 to fallen at frame 50, and from fallen at frame 149 to risen at frame 150, with no transition frames in between.

From eight keyframes to the 50 Hz tracker. The middleware replans every 2 control steps, i.e. every 40 ms. Each plan consists of 8 keyframes spread evenly over the next 0.2 s, and the last keyframe corresponds to the reference at time t + 0.2 s. Joint positions are first clipped to the joint limits, then joint positions and root position are interpolated with cubic splines and root orientation with spherical linear interpolation. The result is a 50 Hz trajectory of 10 frames covering 0.2 s, with joint velocities from finite differences between adjacent frames. From the control step at which a new plan arrives, this trajectory replaces the tracker's reference buffer, and at every step the tracker reads its 10-frame look-ahead window from it to form the reference terms of Sec. 3.3.1.

(3/3)

W3. Ablations of the middleware configuration and of motion sampling

Planning horizon and number of keyframes. We retrained the middleware with different horizons and keyframe counts, each evaluated with the same iFSQ tracker.

Table R4. CR in percent of middleware variantsHorizon (s)KRecovery CR↑Discrete CR↑0.1896.642.90.2, Heracles897.892.00.4895.061.20.8864.715.50.2497.776.00.2698.081.10.21097.378.7\begin{array}{c}\textbf{Table R4. }\text{CR in percent of middleware variants}\cr\begin{array}{|l|c|c|c|}\hline\text{Horizon (s)}&K&\text{Recovery CR}\uparrow&\text{Discrete CR}\uparrow\cr\hline0.1&8&96.6&42.9\cr\text{0.2, Heracles}&8&97.8&\color{red}{92.0}\cr0.4&8&95.0&61.2\cr0.8&8&64.7&15.5\cr0.2&4&97.7&76.0\cr0.2&6&\color{red}{98.0}&81.1\cr0.2&10&97.3&78.7\cr\hline\end{array}\end{array}

On the continuous recovery subset, every horizon between 0.1 and 0.4 s reaches 95.0% to 98.0% CR, and only 0.8 s falls to 64.7%. The discretized subset, with large gaps after each switch, separates the variants clearly. The 0.2 s horizon with K = 8 reaches 92.0%, and every other setting stays at or below 81.1%.

A 0.1 s horizon reaches only 42.9%. The tracker reads a 10-frame look-ahead window at 50 Hz, i.e. 0.2 s, which a 0.2 s plan fills while a 0.1 s plan covers only its first half. Longer horizons lower CR to 61.2% at 0.4 s and 15.5% at 0.8 s, since a longer residual is larger and more varied, which is harder for the same model to fit.

At 0.2 s, recovery CR stays between 97.3% and 98.0% for all keyframe counts, but the discretized subset favors K = 8. K = 6 reaches the highest recovery CR but only 81.1% on discretized references, 10.9 points below K = 8, and K = 4 and K = 10 reach 76.0% and 78.7%. Fewer keyframes describe the transition more coarsely, and with K = 10 the keyframes match the 50 Hz frames in the horizon without improving the plan. Since the discretized subset requires the largest transitions, we choose K = 8.

Adaptive motion sampling. Our sampler, described in App. A, extends the bin-level curriculum of ZEST with continuous bins, kernel-smoothed difficulty propagation and an outlier cap. We retrained the same iFSQ tracker with two alternative samplers, keeping training data, rewards, observations, network and PPO settings unchanged. Uniform sets η=0\eta = 0 in Eq. 19, giving every bin equal probability. Raw failure-proportional samples each continuous bin in proportion to its failure count plus a uniform floor, i.e. our sampler without kernel smoothing and outlier cap. All trackers are evaluated on the 1,000-motion benchmark of Table 1.

Table R5. Tracker trained with different motion samplers, CR in percentSubsetSamplerCR↑Joint (rad)↓Height (m)↓Ori (rad)↓LinVel (m/s)↓OverallUniform79.81.3820.1400.4170.222Raw failure-prop.79.41.8280.1390.4960.240Ours85.21.1090.1050.3800.165DailyUniform96.30.8380.0470.1660.148Raw failure-prop.94.51.2880.0500.2140.237Ours96.30.8020.0390.1850.143RecoveryUniform55.12.4550.2570.8090.381Raw failure-prop.67.52.7820.2030.7950.230Ours92.21.4410.0680.3230.175DiscreteUniform27.42.8500.4621.2000.407Raw failure-prop.20.93.3890.4881.5080.262Ours26.32.2080.4511.3440.258\begin{array}{c}\textbf{Table R5. }\text{Tracker trained with different motion samplers, CR in percent}\cr\begin{array}{|l|l|c|c|c|c|c|}\hline\text{Subset}&\text{Sampler}&\text{CR}\uparrow&\text{Joint (rad)}\downarrow&\text{Height (m)}\downarrow&\text{Ori (rad)}\downarrow&\text{LinVel (m/s)}\downarrow\cr\hline\text{Overall}&\text{Uniform}&79.8&1.382&0.140&0.417&0.222\cr&\text{Raw failure-prop.}&79.4&1.828&0.139&0.496&0.240\cr&\text{Ours}&\color{red}{85.2}&\color{red}{1.109}&\color{red}{0.105}&\color{red}{0.380}&\color{red}{0.165}\cr\hline\text{Daily}&\text{Uniform}&\color{red}{96.3}&0.838&0.047&\color{red}{0.166}&0.148\cr&\text{Raw failure-prop.}&94.5&1.288&0.050&0.214&0.237\cr&\text{Ours}&\color{red}{96.3}&\color{red}{0.802}&\color{red}{0.039}&0.185&\color{red}{0.143}\cr\hline\text{Recovery}&\text{Uniform}&55.1&2.455&0.257&0.809&0.381\cr&\text{Raw failure-prop.}&67.5&2.782&0.203&0.795&0.230\cr&\text{Ours}&\color{red}{92.2}&\color{red}{1.441}&\color{red}{0.068}&\color{red}{0.323}&\color{red}{0.175}\cr\hline\text{Discrete}&\text{Uniform}&\color{red}{27.4}&2.850&0.462&\color{red}{1.200}&0.407\cr&\text{Raw failure-prop.}&20.9&3.389&0.488&1.508&0.262\cr&\text{Ours}&26.3&\color{red}{2.208}&\color{red}{0.451}&1.344&\color{red}{0.258}\cr\hline\end{array}\end{array}

Uniform sampling matches ours on daily motions at 96.3% CR, but its recovery CR drops from 92.2% to 55.1%, with nearly four times the height error, 0.257 m against 0.068 m, and about 2.5 times the orientation error, so the robot frequently fails to get up. Short, difficult segments such as the contact transitions of a get-up are sampled too rarely when all bins weigh the same.

Raw failure-proportional sampling raises recovery CR from 55.1% to 67.5%, confirming that difficulty-based sampling helps. Without smoothing and the outlier cap, however, the distribution collapses, with a single 1 s bin drawing about 75% of all samples. Daily CR then drops to 94.5%, and joint error rises in every subset, to 1.828 rad overall against 1.109 rad.

Our sampler reaches the highest overall CR, is best on all five metrics overall and on the recovery subset, and has the lowest joint error in every subset. On discretized references all three samplers stay between 21% and 27% CR, and in Heracles the middleware raises CR on this subset to 92.0%.

We appreciate the reviewer's careful examination of the experimental design. We hope the additional experiments and details above address the concerns, and we are glad to provide further information in the discussion.

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

This paper introduces HERACLES, which is a state-conditioned generative middleware. It provides a bridge between a reference motion and a humanoid controller. Instead of always receiving the original reference, the controller receives an adapted motion trajectory. This trajectory is generated from the current state of the robot and the desired reference using a conditioned flow matching model. The middleware aims to leave the reference to a great extent unchanged when the robot is close to the reference, while generating intermediate movements when substantial deviations occur.

Critical strengths
  • The separation of the trajectory generator from the controller is useful, as the middleware provides an adapted reference to an already trained tracker.
  • The results also appear to be supporting this design in the settings of Overall and Discrete, in comparison to the models without this, where they are underperforming.
Critical weaknesses
  • The results are encouraging, but these improvements are not consistent across motion subsets and metrics. While HERACLES performs strongly in Overall and Discrete settings, it reduces completion on daily motions, and worsens joint-position, height, and linear velocity errors relative to the standalone iFSQ tracker. This is problematic despite the improvements in orientation accuracy. Additionally, even on Discrete setting, where the gains are strong, the linear velocity error increases. These trade-offs require clearer analysis and discussion, especially in relation to the claim that the middleware preserves nominal tracking fidelity.

  • The paper performs several component ablations regarding directional initialization, noisy state augmentation, and kinematics-aware weighting. However, these are mainly reported at the aggregated (Overall) benchmark level, and they do not directly establish the claimed near identity behaviour or explain the degradation on the Daily motion subset. Therefore, reporting the existing ablations by motion subset, along with a direct analysis of reference modifications when state reference discrepancies are small, would strengthen this claim. A comparison against a simple interpolation based planner using the same tracker would also help to distinguish the contribution of the learned generative model from the benefit of supplying intermediate reference poses.

Questions for authors

Q1: The results in Table 1 show a lower Completion Rate and higher errors in several tracking metrics on the Daily motion subset compared with the standalone iFSQ tracker. Could the authors please quantify how much the middleware modifies the original reference trajectory when the state of robot is already close to it? Does this mean that larger modifications are associated with higher tracking errors or lower completion rates on these motions?

Q2: Could the authors please compare HERACLES against a simple interpolation based reference planner, while keeping the tracker, planning horizon, and replanning frequency unchanged? This comparison would help to clarify the contribution of the learned generative model beyond simply supplying intermediate reference poses, especially in the case of the Discrete reference subset.

2 comments

(1/2)

We thank the reviewer for the thoughtful review and the concrete suggestions. To answer the two questions, we ran two additional analyses, reported in two comments. The first measures how much the middleware modifies the reference on daily motions, and the second compares Heracles with an interpolation-based planner under the same tracker, horizon and replanning frequency.

Q1. Modification of the reference and its relation to tracking error and CR

The differences between Heracles and standalone iFSQ on the daily subset come from a tracking target that only Heracles has. The 38-D middleware is conditioned on both the current and the reference root position, so each plan also steers the robot toward the reference root. iFSQ alone observes only root velocities and orientation error, as defined in Sec. 3.3.1, and cannot correct horizontal drift. No metric in Table 1 involves horizontal root position, so these adjustments appear there only as lower CR and higher joint, height and velocity errors.

To quantify the modification, at every control step we compare the final keyframe of the current plan, 0.2 s ahead, with the original reference at that time. We use the final keyframe because the first one is anchored to the current state and reflects the robot's deviation, not the middleware's change. Joint modification is the mean absolute difference over the 29 joints, and root modification is the horizontal and vertical distance. We use all control steps of 38-D Heracles on the daily subset before the first failure of each clip, about 296k steps, split into quintiles of the current joint tracking error.

Table R1. Median modification by current tracking error, daily, joints in rad/joint, root in mQuintileTracking errorJointRoot xyRoot height1, closest0.0700.0140.0050.00320.0990.0230.0080.00530.1260.0480.0170.00840.1640.0810.0390.0125, farthest0.3080.1710.0820.022\begin{array}{c}\textbf{Table R1. }\text{Median modification by current tracking error, daily, joints in rad/joint, root in m}\cr\begin{array}{|c|c|c|c|c|}\hline\text{Quintile}&\text{Tracking error}&\text{Joint}&\text{Root xy}&\text{Root height}\cr\hline\text{1, closest}&0.070&0.014&0.005&0.003\cr 2&0.099&0.023&0.008&0.005\cr 3&0.126&0.048&0.017&0.008\cr 4&0.164&0.081&0.039&0.012\cr\text{5, farthest}&0.308&0.171&0.082&0.022\cr\hline\end{array}\end{array}

When the robot is closest to the reference, the middleware changes each joint by 0.014 rad, about 0.8 degrees, and moves the root by about 5 mm horizontally and 3 mm vertically, so the reference passes through essentially unchanged.

Table R2. Modification and Spearman correlation with per-clip joint error and CR, 700 daily clipsMedianMean95th pct.ρ, joint errorρ, CRJoint (rad/joint)0.0470.0770.252+0.84−0.50Root xy (m)0.0150.1090.511+0.71−0.47Root height (m)0.0070.0160.055+0.74−0.52\begin{array}{c}\textbf{Table R2. }\text{Modification and Spearman correlation with per-clip joint error and CR, 700 daily clips}\cr\begin{array}{|l|c|c|c|c|c|}\hline&\text{Median}&\text{Mean}&\text{95th pct.}&\rho\text{, joint error}&\rho\text{, CR}\cr\hline\text{Joint (rad/joint)}&0.047&0.077&0.252&+0.84&-0.50\cr\text{Root xy (m)}&0.015&0.109&0.511&+0.71&-0.47\cr\text{Root height (m)}&0.007&0.016&0.055&+0.74&-0.52\cr\hline\end{array}\end{array}

Clips with larger modifications do have higher tracking error and lower CR. This correlation arises because the middleware responds to deviations that have occurred. It plans from the robot's current state, so the further the robot has drifted, the larger the change it makes to bring it back, and in Table R1 the modification grows steadily with the current tracking error.

These modifications move the robot back toward the reference.

Table R3. Root horizontal error (m) over time, 190 daily clips of at least 10 sTime (s)246810iFSQ0.0560.1050.1730.2100.226Heracles0.0450.0930.1170.1270.094\begin{array}{c}\textbf{Table R3. }\text{Root horizontal error (m) over time, 190 daily clips of at least 10 s}\cr\begin{array}{|l|c|c|c|c|c|}\hline\text{Time (s)}&2&4&6&8&10\cr\hline\text{iFSQ}&0.056&0.105&0.173&0.210&0.226\cr\text{Heracles}&\color{red}{0.045}&\color{red}{0.093}&\color{red}{0.117}&\color{red}{0.127}&\color{red}{0.094}\cr\hline\end{array}\end{array}

Without the middleware, drift accumulates to 0.226 m after 10 s. With it, drift stays bounded and decreases after 8 s, ending 58% lower, as also shown in a newly added supplementary video.

To confirm that root tracking explains the daily differences, we replace the current root position in the middleware input with the reference root position, with weights, tracker and inference unchanged, so no horizontal correction is planned. Results are mean ±\pm std over three middleware seeds.

Table R4. Daily subset, CR in percentCR↑Joint (rad)↓Height (m)↓Ori (rad)↓LinVel (m/s)↓iFSQ96.30.8020.0390.1850.143Heracles, 38-D93.70.8950.0500.1420.152Heracles, 38-D, ref. root input95.3±0.20.796±0.0040.045±0.0010.141±0.0030.102±0.001\begin{array}{c}\textbf{Table R4. }\text{Daily subset, CR in percent}\cr\begin{array}{|l|c|c|c|c|c|}\hline&\text{CR}\uparrow&\text{Joint (rad)}\downarrow&\text{Height (m)}\downarrow&\text{Ori (rad)}\downarrow&\text{LinVel (m/s)}\downarrow\cr\hline\text{iFSQ}&\color{red}{96.3}&0.802&\color{red}{0.039}&0.185&0.143\cr\text{Heracles, 38-D}&93.7&0.895&0.050&0.142&0.152\cr\text{Heracles, 38-D, ref. root input}&95.3\pm0.2&\color{red}{0.796\pm0.004}&0.045\pm0.001&\color{red}{0.141\pm0.003}&\color{red}{0.102\pm0.001}\cr\hline\end{array}\end{array}

With horizontal correction removed, CR rises from 93.7% to 95.3% and joint error drops to 0.796 rad, below iFSQ, as do orientation and velocity errors. The daily differences in Table 1 therefore come from following the global root position rather than from the generative middleware itself.

The discretized subset and Q2 follow in the next comment.

(2/2)

Linear velocity error on the discretized subset

The 150 discretized references are built from fall-and-get-up motions in the proprietary motion-capture collection described in Sec. 4.1, with no overlap with the daily or recovery subsets. Each reference lasts 550 frames at 50 fps and consists of three static poses, each a single stable frame of the same capture, held for frames 0-49, 50-149 and 150-549, i.e. 1 s, 2 s and 8 s. Adjacent poses differ in pelvis height by at least 0.30 m, so every switch goes between standing and lying. Linear velocity error measures how far the robot's root linear velocity deviates from the root linear velocity of the reference motion, and every metric is computed against the original reference, as stated in App. D. Since every segment is static, the reference root velocity is zero throughout, and any root motion of the robot counts as velocity error.

A policy that falls at the first switch and stays down barely moves and incurs little velocity error, yet satisfies the CR criterion during the standing and fallen segments, about 3/11 or 27% of the frames. This matches the CR of the trackers that never rise, namely MLP 28.6%, SONIC 28.4%, iFSQ 26.3%, VAE 23.4% and Transformer 20.6%, and the Transformer, with the lowest CR, also has the lowest velocity error, 0.158 m/s. Heracles is the only system that executes both transitions, which produces root velocity where the reference has none. Its height and orientation errors of 0.055 m and 0.159 rad, which measure whether the commanded poses are reached, are the lowest on the subset.

Taken together, when the robot tracks the reference closely the middleware changes it very little, the daily-subset differences come from following the global root position, and the higher velocity error on discretized references results from executing the commanded transitions against a reference whose root velocity is zero.

Q2. Comparison with an interpolation-based reference planner

The interpolation baseline replaces only the generative step of the middleware. At each planning cycle it computes the eight keyframes directly, interpolating the 29 joint angles and the root position linearly from the current state to the target and the root orientation by spherical linear interpolation. As in Heracles, the first keyframe equals the current state, and the root horizontal position is expressed relative to the current root. Everything else is shared with Heracles, namely the 38-D state, the 0.2 s horizon with K = 8, replanning at 25 Hz, the target passed to the planner, the conversion of keyframes into the 50 Hz reference, the iFSQ tracker checkpoint, the evaluation code and all metrics.

Table R5. Discretized subset with CR in percent, and recovery SR on the 40 fallen configurations of Table 10CR↑Joint (rad)↓Height (m)↓Ori (rad)↓LinVel (m/s)↓SR↑iFSQ, no planner26.32.2080.4511.3440.2580/40iFSQ + interpolation planner15.12.7780.5181.4530.2010/40Heracles, iFSQ + flow matching92.01.1810.0550.1590.31437/40\begin{array}{c}\textbf{Table R5. }\text{Discretized subset with CR in percent, and recovery SR on the 40 fallen configurations of Table 10}\cr\begin{array}{|l|c|c|c|c|c|c|}\hline&\text{CR}\uparrow&\text{Joint (rad)}\downarrow&\text{Height (m)}\downarrow&\text{Ori (rad)}\downarrow&\text{LinVel (m/s)}\downarrow&\text{SR}\uparrow\cr\hline\text{iFSQ, no planner}&26.3&2.208&0.451&1.344&0.258&0/40\cr\text{iFSQ + interpolation planner}&15.1&2.778&0.518&1.453&0.201&0/40\cr\text{Heracles, iFSQ + flow matching}&\color{red}{92.0}&\color{red}{1.181}&\color{red}{0.055}&\color{red}{0.159}&0.314&\color{red}{37/40}\cr\hline\end{array}\end{array}

With interpolated poses, CR on the discretized subset falls from 26.3% to 15.1%, below the tracker without a planner, and none of the 40 fallen configurations is recovered, while Heracles reaches 92.0% and 37/40. Two properties of linear interpolation explain this.

  • The full gap is compressed into one planning horizon. Each interpolated plan spans the entire distance to the target within 0.2 s. Successive discretized poses differ in pelvis height by 0.37 to 0.90 m, as reported in Sec. 4.1, so each plan asks the tracker to cover such a change in 0.2 s. The flow matching middleware predicts only the motion increment over the next 0.2 s, as described in Sec. 3.2, so its plan stays bounded however far the target lies.
  • Linear interpolation does not follow the support phases of a get-up. Rising from the ground passes through support phases such as rolling, propping on the arms and kneeling, whose configurations lie far from the straight line between lying and standing and sometimes move away from the target. A linear plan skips them and gives the tracker poses that no contact sequence supports, whereas the middleware, trained on recorded human transitions, follows the contact sequences present in the data.

Videos of the interpolation planner on 10 randomly selected fallen configurations from Table 10 are in the updated supplementary material, and this comparison is reported in App. I of the revised paper.

We thank the reviewer again for the constructive suggestions, which led to the additional analyses above. We hope these results address the concerns and are glad to provide further details in the discussion.

Reviewer_PR8vOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Heracles, a flow-matching middleware that generates short motion transitions between a humanoid’s current state and its reference motion. It operates above a low-level tracker, allowing the robot to handle large deviations without explicitly switching to a recovery policy. The authors evaluate the approach on 1,000 held-out sequences, demonstrate transfer to SONIC, and deploy the system on a Unitree G1.

Critical strengths
  1. The middleware addresses a meaningful problem with a practical, modular solution. On discontinuous references, adding it to iFSQ improves CR from 26.3% to 92.0%. Autonomous recovery also improves from 0/40 to 37/40 configurations. These are substantial gains that directly support the main contribution.

  2. The SONIC transfer experiment is particularly useful because it separates the middleware’s benefit from improvements to the proposed tracker. Recovery improves from 0/40 to 18/40 without changing SONIC’s weights. Onboard execution and physical demonstrations further support the approach’s practical feasibility.

Critical weaknesses
  1. The claim that nominal tracking fidelity is preserved is too strong. On daily motions, CR decreases from 96.3% to 93.7%, while joint error increases from 0.802 to 0.895 rad after adding the middleware. This suggests a trade-off between nominal precision and recovery capability that deserves explicit analysis.

  2. The comparisons establish the benefit of the middleware more clearly than the benefit of its generative formulation. A deterministic trajectory predictor trained on the same data would help isolate the contribution of flow matching. Furthermore, framewise CR and three hardware recovery trials provide limited evidence for the broader robustness claims.

Questions for authors
  1. What causes the reduction in nominal tracking performance, and is this trend consistent across training seeds?

  2. Have you compared flow matching with a deterministic state-conditioned trajectory predictor using the same training data and tracker?

  3. Were hardware recoveries tested using only a static standing reference, rather than a live performer? If so, please report the number of trials and success rate.

1 comment

We thank the reviewer for the careful reading and the precise questions. To answer them, we ran three new experiments, reported below.

Q1. Cause of the reduction in nominal tracking and its consistency across seeds

Beyond the quantities reported in Table 1, Heracles also tracks the global root position of the reference. Its planner receives both root positions and steers the robot toward the reference root, whereas the iFSQ tracker observes only root velocities and orientation error, Sec. 3.3.1 and cannot correct horizontal drift. In Table 1, this extra root-position tracking appears as lower CR and higher joint, height and velocity errors than iFSQ.

Table R1. Root horizontal error (m), 190 daily clips of at least 10 sTime (s)246810iFSQ0.0560.1050.1730.2100.226Heracles0.0450.0930.1170.1270.094\begin{array}{c}\textbf{Table R1. }\text{Root horizontal error (m), 190 daily clips of at least 10 s}\cr\begin{array}{|l|c|c|c|c|c|}\hline\text{Time (s)}&2&4&6&8&10\cr\hline\text{iFSQ}&0.056&0.105&0.173&0.210&0.226\cr\text{Heracles}&\color{red}{0.045}&\color{red}{0.093}&\color{red}{0.117}&\color{red}{0.127}&\color{red}{0.094}\cr\hline\end{array}\end{array}

iFSQ drifts monotonically to 0.226 m, while Heracles stays bounded and ends 58% lower, as also shown in a newly added supplementary video.

To verify this, we feed the middleware the reference root position in place of the measured one, with weights, tracker and inference unchanged. We retrained the middleware with two more seeds and report mean ±\pm std over three seeds.

Table R2. Daily subset, CR in percentCR↑Joint (rad)↓Height (m)↓Ori (rad)↓LinVel (m/s)↓iFSQ96.30.8020.0390.1850.143Heracles93.70.8950.0500.1420.152Heracles, ref. root input95.3±0.20.796±0.0040.045±0.0010.141±0.0030.102±0.001\begin{array}{c}\textbf{Table R2. }\text{Daily subset, CR in percent}\cr\begin{array}{|l|c|c|c|c|c|}\hline&\text{CR}\uparrow&\text{Joint (rad)}\downarrow&\text{Height (m)}\downarrow&\text{Ori (rad)}\downarrow&\text{LinVel (m/s)}\downarrow\cr\hline\text{iFSQ}&\color{red}{96.3}&0.802&\color{red}{0.039}&0.185&0.143\cr\text{Heracles}&93.7&0.895&0.050&0.142&0.152\cr\text{Heracles, ref. root input}&95.3\pm0.2&\color{red}{0.796\pm0.004}&0.045\pm0.001&\color{red}{0.141\pm0.003}&\color{red}{0.102\pm0.001}\cr\hline\end{array}\end{array}

Without root-position tracking, CR rises from 93.7% to 95.3% and joint error falls to 0.796 rad, below iFSQ, as do orientation and velocity errors. Seed spread is at most 0.2 points of CR and 0.004 on every error. The difference in Table 1 thus comes from tracking the global root position, not from the generative middleware itself.

Q2. Flow matching compared with a deterministic state-conditioned predictor

Both middlewares share the training data and augmentation, conditioning, residual parameterization, kinematics-aware weighting, AdaLN Transformer, optimizer and a comparable training budget. The deterministic predictor has no noise input or timestep embedding, decodes eight learnable query tokens in one forward pass, and is trained with weighted MSE. Flow matching uses 5 Euler steps with a warm start at tstart=0.9t_{\text{start}} = 0.9, as in the paper. Both drive the same iFSQ checkpoint.

Table R3. Recovery on the 40 fallen configurations of Table 10SystemSR↑iFSQ, no middleware0/40iFSQ + deterministic predictor22/40Heracles, iFSQ + flow matching37/40\begin{array}{c}\textbf{Table R3. }\text{Recovery on the 40 fallen configurations of Table 10}\cr\begin{array}{|l|c|}\hline\text{System}&\text{SR}\uparrow\cr\hline\text{iFSQ, no middleware}&0/40\cr\text{iFSQ + deterministic predictor}&22/40\cr\text{Heracles, iFSQ + flow matching}&\color{red}{37/40}\cr\hline\end{array}\end{array}

The deterministic predictor reaches 22/40 where the tracker alone reaches 0/40, so a state-conditioned middleware alone gives partial ability to reach a distant reference. Flow matching recovers 15 more, for two reasons.

  • Multimodality. A fallen state admits several ways to rise, such as rolling prone or sitting up first. MSE regression returns their conditional mean, which follows none of the incompatible strategies and shrinks the residual toward zero. Flow matching models the full conditional distribution, so each plan commits to one strategy in the data.
  • Fresh sampling at every replanning cycle. The ODE is deterministic, but each cycle starts from new Gaussian noise, Eq. 10. When progress is difficult, successive plans explore other ways forward, whereas the deterministic predictor repeats nearly the same plan and the robot stays stuck.

Videos on 10 random configurations are in the updated supplementary material, and this comparison is in App. I of the revised paper.

Q3. Hardware recovery with only a static standing reference

In the third fall-and-recovery row of Fig. 3, the performer stands up from the third frame while the robot is still on the ground, so the rise from that point is planned entirely by the middleware.

We also ran a dedicated hardware test on 17 supine, prone and lateral fallen poses from the validation set. After tracking a reference into the pose, the robot receives only one static standing frame, the same for all trials, and the middleware alone plans the entire rise. The robot stood up in all 17 consecutive trials. The recording and its simulation replay are in the updated supplementary material.

We are grateful for the suggestions, which directly led to the new experiments above, and are glad to provide further details in the discussion.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Revisions to the Paper and Supplementary Material

We thank all reviewers and the area chair for their time and constructive feedback. Following the discussion, we have uploaded a revised PDF and an updated supplementary material. The changes are listed below, together with the reviewer questions they address.

Revised PDF

The main text is unchanged. All changes are confined to the appendix, and the numbering of the existing sections, tables and equations is preserved, so references in the original submission and in our responses remain valid.

  • New App. I, Alternative Middleware Formulations. To isolate the contribution of the generative formulation, we replace the flow matching network with two alternatives while keeping every other component unchanged, namely a deterministic state-conditioned predictor trained with weighted MSE and a planner that linearly interpolates from the current state to the target. Table 12 reports autonomous recovery on the 40 fallen configurations of Table 10, and Table 13 reports the interpolation planner on the recovery and discretized subsets. This appendix addresses Q2 of Reviewer PR8v, Q2 of Reviewer bBum and W1 of Reviewer Qp3B.
  • App. E. One sentence is added to point to the new App. I.
  • Limitations and Future Work. This section moves from App. I to App. J because of the new appendix. Its content is unchanged.

Updated supplementary material

The four videos of the original submission are unchanged. We added four videos, numbered in the order in which they are referenced in our responses.

  • rebuttal_1_global_root_drift_correction.mp4 compares the root trajectories of the standalone iFSQ tracker and Heracles on daily motions and shows how the 38-D middleware corrects accumulated horizontal drift. It accompanies Q1 of Reviewer PR8v and Q1 of Reviewer bBum.
  • rebuttal_2_flow_matching_vs_deterministic_planner.mp4 shows the flow matching middleware and the deterministic predictor on 10 randomly selected fallen configurations of Table 10. It accompanies Q2 of Reviewer PR8v.
  • rebuttal_3_hardware_17_consecutive_getups.mp4 shows 17 consecutive get-ups on the physical robot from supine, prone and lateral fallen poses, each commanded only by a single static standing frame, together with a simulation replay of the recorded log. It accompanies Q3 of Reviewer PR8v.
  • rebuttal_4_flow_matching_vs_linear_interpolation.mp4 shows the flow matching middleware and the linear interpolation planner on 10 randomly selected fallen configurations of Table 10. It accompanies Q2 of Reviewer bBum and W1 of Reviewer Qp3B.

Further new results, including the analysis of nominal tracking, the ablations of the planning horizon, the number of keyframes and the motion sampler, and the construction of the discretized references, are reported in the individual responses. We are glad to provide further details during the discussion.

We thank the reviewers for their careful reading and constructive questions.

Consensus across the reviews. The reviews agree that the middleware is a practical, modular solution that attaches to existing trackers without retraining (PR8v, bBum, Qp3B), that it brings substantial gains when the reference is far from the robot, raising discretized CR from 26.3\color{red}{\textbf{26.3}}% to 92.0\color{red}{\textbf{92.0}}% and recovery from 0/40\color{red}{\textbf{0/40}} to 37/40\color{red}{\textbf{37/40}} (PR8v, bBum, Qp3B), and that its transfer to SONIC separates its contribution from that of our tracker (PR8v, Qp3B).

Questions addressed in our responses.

1. Daily motions (PR8v, bBum). Near the reference, the middleware changes it by only 0.014\color{red}{\textbf{0.014}} rad per joint and about 5\color{red}{\textbf{5}} mm in root position. The daily-subset differences come from correcting horizontal drift, which Table 1 does not measure. Without this correction, joint error falls below the standalone tracker, consistently across three seeds.

2. Generative formulation (PR8v, bBum, Qp3B). With the same tracker, a deterministic predictor recovers 22/40\color{red}{\textbf{22/40}} and linear interpolation 0/40\color{red}{\textbf{0/40}}, against 37/40\color{red}{\textbf{37/40}} for flow matching.

3. Hardware (PR8v). The robot stood up in 17\color{red}{\textbf{17}} of 17\color{red}{\textbf{17}} consecutive trials from a single static standing frame.

4. Tracker choice, ablations and details (Qp3B). We explain the VQ-VAE comparison, add ablations of horizon, keyframes and motion sampling, and document the discretized references and the planner-tracker interface.

We hope our responses have addressed your concerns, and we would be grateful for your thoughts before the discussion period ends on November 18\color{red}{\textbf{November 18}}.

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

Related papers

Discussion (0)

Sign in to comment.