PaperlayerView exact revision on arXiv

Source-linked AI summary

Whole-Body Motion Guided Reference Generation for Uncertainty-Aware Compliant Assistance

Muyuan Ma, Long Cheng, Lijun Han, Haotian Zhai, Xinpan Meng, Xiuze Xia, Houcheng Li, Zeyu Liu

arXiv:2610.31302v1cs.RO

TL;DR

Reliable exoskeleton assistance is difficult when motion and interaction conditions vary and reference generation lacks richer multimodal information. This paper couples whole-body-motion-guided trajectory generation with uncertainty-aware variable admittance control. Experiments report lower muscle activation than baselines and lower metabolic energy expenditure during repetitive box lifting than without assistance.

  • Problem

    Exoskeleton assistance must work across varying users and tasks, but many reference methods rely on limited signals and existing controllers provide little information about reference reliability.

    Source: §I · pp. 1–2

  • Method

    The framework generates trajectories from whole-body motion and exoskeleton proprioception, uses task descriptions as auxiliary cues, and feeds sampled-trajectory uncertainty into variable admittance stiffness regulation.

    Source: §I · p. 2

  • Results

    The framework significantly reduces muscle activation compared with baselines and decreases metabolic energy expenditure during repetitive box lifting.

    Source: §V · p. 16

  • Takeaways & Limitations

    Uncertainty-aware stiffness regulation balances tracking the generated reference with compliant interaction under stability constraints.

    Source: §I · p. 2, §III · p. 8

Abstract

from arXiv · show

Exoskeletons have shown great potential for augmenting human physical capabilities, yet generating reliable assistance remains challenging because human motion patterns and interaction requirements vary substantially in practice. To address this problem, a dual-loop assistance strategy is proposed to couple multimodal exoskeleton trajectory generation with variable admittance control, where trajectory variance links the outer-loop generator and the inner-loop controller. The generator synthesizes future exoskeleton trajectories through denoising, primarily using whole-body human motion and exoskeleton proprioception, with task descriptions as auxiliary semantic cues. To reduce cross-modal representation mismatch, a three-stage training pipeline is introduced to progressively optimize the human motion encoder, the exoskeleton proprioception encoder, and the trajectory generator. In the inner loop, multiple future trajectories are sampled from the generator to estimate the trajectory mean and variance. The mean is used as the control reference, while the variance provides an uncertainty measure for prediction reliability. This measure is further incorporated into a stiffness regulation law, enabling compliant assistance with input-to-state stability under bounded lumped interaction input. Experimental results demonstrate that the proposed method significantly reduces muscle activation levels compared with standard baselines, while also decreasing users' metabolic energy expenditure relative to the unassisted condition.

I. INTRODUCTION

Reliable assistance is challenging because motion and interaction vary across users and tasks, while existing references may lack rich multimodal context and controllers have limited information about reference reliability. The paper couples multimodal trajectory generation with uncertainty-aware admittance control, using whole-body motion, exoskeleton proprioception, and task cues to generate references.

I. INTRODUCTION

  • Assistance must remain effective and comfortable across varying users and tasks, but existing reference generation and feedback control leave gaps in multimodal context and trajectory reliability.Existing controllers largely respond to interaction forces or tracking errors, which signal disagreement after it occurs but offer limited information about reference reliability.Source: §I · pp. 1–2
  • The dual-loop framework couples multimodal trajectory generation and variable admittance control through trajectory uncertainty, which regulates stiffness alongside the generated reference.The mean of sampled trajectories supplies the reference; higher uncertainty lowers stiffness, while lower uncertainty raises it, subject to stability constraints.Source: §I · p. 2

II. MULTIMODAL EXOSKELETON TRAJECTORY GENERATION

  • Whole-body human motion and exoskeleton proprioception provide time-varying physical conditions, while task descriptions provide static semantic cues for the intended assistance scenario.The paper describes whole-body motion as adding global context beyond local exoskeleton proprioception.Source: §II · p. 2, §I · p. 2

A. Generator Architecture

  • The generator encodes human motion, task descriptions, and exoskeleton proprioception into shared-dimension tokens that condition a Transformer decoder denoising future trajectories.Trajectory tokens serve as decoder queries, while multimodal tokens form conditional memory for learning a conditional velocity field.Source: A. Generator Architecture · p. 3

B. Three-Stage Training Pipeline

The training pipeline progressively adapts from large-scale human motion data to simulated interaction data and then to real exoskeleton trajectories. It transfers pretrained motion representations, aligns proprioception in simulation, and uses parameter-efficient decoder adaptation for limited real data.

B. Three-Stage Training Pipeline

  • The three-stage pipeline moves from large-scale human-motion pretraining, through simulated multimodal alignment, to adaptation on real exoskeleton trajectories.Dataset scale decreases as task relevance increases across stages.Source: B. Three-Stage Training Pipeline · p. 4

Stage 1: Human Motion Pretraining

  • Stage 1 pretrains the human motion encoder on paired text-motion data, then transfers it as a motion prior and discards the auxiliary Transformer decoder.HumanML3D contains 14,616 motion sequences and 44,970 natural-language descriptions.Source: 1) Stage 1: Human Motion Pretraining · p. 4

Stage 2: Simulated Proprioception Alignment

  • Stage 2 uses simulated human–exoskeleton interaction data to align human motion, task descriptions, and proprioception before training trajectory generation.The simulation randomizes body mass, assistance ratio, and assistance side, producing data for 10 motion tasks.Source: 2) Stage 2: Simulated Proprioception… · pp. 4–5

Stage 3: Real Trajectory Adaptation

  • Stage 3 adapts the simulation-trained generator to real sensing and interaction conditions using data from three users performing 13 tasks.The real dataset provides 156 training trajectories; the final repetition for each task is reserved for evaluation, and only decoder LoRA parameters are optimized.Source: 3) Stage 3: Real Trajectory Adaptation · p. 5

III. UNCERTAINTY-AWARE VARIABLE ADMITTANCE CONTROL

The controller uses sampled trajectory uncertainty to adjust interaction stiffness while tracking the sample-mean reference. Its closed-loop tracking error is input-to-state stable under bounded interaction input.

III. UNCERTAINTY-AWARE VARIABLE ADMITTANCE CONTROL

  • The sample mean of N generated trajectories sets the reference, while their averaged sample variance measures uncertainty for stiffness regulation.The generator is repeatedly sampled at each outer-loop trajectory update using the same history, task description, and proprioceptive states.Source: §III · p. 6
  • The admittance model regulates motion deviation from the generated reference using virtual inertia, damping, and time-varying interaction stiffness.The position, velocity, and acceleration deviations are measured relative to the desired joint position from the generated reference.Source: §III · p. 6
  • Increasing trajectory uncertainty lowers interaction stiffness toward Klow for greater compliance; decreasing uncertainty raises it toward Kup for stronger reference tracking.The regulation uses a filtered, normalized trajectory standard deviation and bounded uncertainty-variation signal.Source: §III · pp. 6–7
  • Bounded human–exoskeleton interaction force and gravity-compensation error imply bounded tracking errors; with zero lumped input, the error converges exponentially.The closed-loop error system is input-to-state stable with respect to the lumped input d(t).Source: §III · pp. 7–8

IV. EXPERIMENTAL VALIDATION

The evaluation has two groups: ablations of training and input choices, and human-subject comparisons of the proposed generator and controller against representative baselines.

IV. EXPERIMENTAL VALIDATION

  • Human-subject comparisons assess trajectory generation, interaction control, and overall assistance against representative baselines.This is the second experimental group.Source: §IV · p. 8
  • Ablation studies examine contributions from the three-stage training pipeline and multimodal conditional inputs.These form the first experimental group.Source: §IV · p. 8
  • The experiments evaluate the proposed uncertainty-aware multimodal assistance framework through both ablations and human-subject comparisons.The two groups cover component contributions as well as comparative assistance performance.Source: §IV · p. 8

A. Ablation Studies

The ablations evaluate future exoskeleton trajectory prediction on the real-data evaluation set using three error metrics across 13 tasks.

A. Ablation Studies

  • Predicted future exoskeleton trajectories are compared with the true trajectories on the real exoskeleton evaluation set.The set covers 13 tasks, including object manipulation, gesture, and weightlifting.Source: A. Ablation Studies · p. 8
  • The evaluation set spans 13 real exoskeleton tasks, including object manipulation, gesture, and weightlifting.The task list is summarized in Table I.Source: A. Ablation Studies · p. 8
  • Mean absolute error, mean squared error, and root mean squared error quantify trajectory prediction errors.The metrics assess prediction under different training schemes and multimodal conditional inputs.Source: A. Ablation Studies · p. 9

1) Ablation on the Three-Stage Training Pipeline:

The complete three-stage training pipeline yields the lowest errors, while simulated interaction training and large-scale human-motion pretraining each contribute improvements over real-data-only adaptation.

1) Ablation on the Three-Stage Training Pipeline:

  • Using only real-data adaptation raises exoskeleton MAE by 35.59%, MSE by 47.90%, and RMSE by 25.54% versus the complete pipeline.The complete three-stage pipeline achieves the lowest errors among the training schemes.Source: 1) Ablation on the Three-Stage… · p. 9
  • Adding simulated human–exoskeleton interaction training reduces prediction errors versus real-data-only adaptation.The passage attributes useful exoskeleton-specific motion knowledge to simulated proprioceptive alignment.Source: 1) Ablation on the Three-Stage… · p. 9
  • Adding large-scale human-motion pretraining reduces exoskeleton MAE by 9.21%, MSE by 12.72%, and RMSE by 5.93% beyond simulated training.The passage identifies this pretraining as a beneficial motion prior for real exoskeleton trajectory generation.Source: 1) Ablation on the Three-Stage… · p. 9

2) Ablation on Multimodal Conditional Inputs:

Removing human motion causes the largest prediction-error increase, while task descriptions provide a distinct semantic cue when identical physical observations correspond to different intended trajectories. Attention patterns and a Task 5 example further illustrate the complementary roles of physical modalities.

2) Ablation on Multimodal Conditional Inputs:

  • Removing human motion increases exoskeleton MAE, MSE, and RMSE by 16.47%, 29.74%, and 12.26%, respectively, the largest degradation among the tested input removals.Full multimodal inputs achieve the lowest errors; removing task descriptions causes only a marginal increase.Source: 2) Ablation on Multimodal Conditional… · p. 9, Table III · p. 10
  • On ambiguous-intention samples, mismatched task descriptions raise RMSE from 0.503 ± 0.221 rad to 0.850 ± 0.304 rad, with Pcor<mis reaching 79.49%.Paired samples have identical historical human motion and exoskeleton proprioception but different descriptions and future trajectories.Source: 2) Ablation on Multimodal Conditional… · p. 9
  • Across most tasks, exoskeleton proprioception receives relatively high attention, human-motion attention varies by task, and task-description attention is lower and relatively stable.The attention visualization covers 13 real exoskeleton tasks.Source: 2) Ablation on Multimodal Conditional… · p. 10, Figure 7 · p. 10
  • In Task 5, similar observed exoskeleton states accompany different human-motion patterns across movement onset and object approach, and predictions match the corresponding future trajectories.The example shows the subject keeping the arm approximately vertical during approach, then flexing the elbow after contact to lift the object.Source: 2) Ablation on Multimodal Conditional… · p. 10, Figure 8 · p. 10

B. Comparative Experiments

Comparative experiments evaluate trajectory generation, uncertainty-guided interaction control, and overall assistance in seven previously unseen subjects. The test subjects had no prior exoskeleton experience and were excluded from training, model selection, and tuning.

B. Comparative Experiments

  • The framework is evaluated against representative baselines in trajectory generation, interaction control, and overall assistance using seven previously unseen subjects.The subjects had no prior exoskeleton experience and were not involved in training, model selection, or tuning.Source: B. Comparative Experiments · pp. 10–11
  • Trajectory generation is compared with the intention predictor in, while uncertainty-guided stiffness regulation is compared with the tracking-error-based compliance method in.For overall assistance, the complete framework is compared with the unassisted condition.Source: B. Comparative Experiments · p. 11
  • The setup uses an RGB-D camera to estimate human-motion keypoints and a USB-CAN adapter to transmit motor commands to the elbow exoskeleton.Interaction signals are acquired through a USB-ADC module; the host computer runs the generator and control algorithm.Source: B. Comparative Experiments · p. 11, Figure 9 · p. 11

1) Comparison of Multimodal Trajectory Generation:

The multimodal generator produces more accurate reference trajectories than the intention-prediction baseline under identical control settings, and it generally reduces muscle activation. The reduction is most evident in the biceps brachii without compensatory increases in the other measured muscles.

1) Comparison of Multimodal Trajectory Generation:

  • 0.257 ± 0.095 rad versus 0.376 ± 0.208 rad: the proposed generator reduces trajectory tracking error by 31.59% compared with.Both methods use the same variable admittance controller and parameter settings, so the RMSE primarily reflects reference-trajectory quality.Source: 1) Comparison of Multimodal Trajectory… · p. 11
  • The proposed method produces lower biceps brachii activation in all reported tasks, with statistically significant reductions in nearly all cases.The biceps brachii is the primary agonist for elbow flexion; brachioradialis and triceps activation are also lower in most tasks.Source: 1) Comparison of Multimodal Trajectory… · pp. 12–13, Table IV · p. 13
  • Reduced biceps brachii activation is not accompanied by compensatory increases in the auxiliary flexor or antagonist muscles.The reported pattern is lower primary elbow-flexion effort while maintaining or lowering activation in the other related muscles.Source: 1) Comparison of Multimodal Trajectory… · p. 13

2) Comparison of Variable Admittance Control:

Trajectory variance is positively associated with tracking error across tasks and subjects, supporting its use as a reliability indicator for stiffness regulation. Compared with the tracking-error-based controller, the proposed variable admittance controller lowers interaction force under all tested loads and lowers muscle activation across measured muscles and load conditions.

2) Comparison of Variable Admittance Control:

  • r = 0.6558 and ρ = 0.7198: trajectory variance is positively correlated with tracking error across tasks and subjects.Tracking error rises from 0.0726 ± 0.0877 in the low-variance group to 0.4229 ± 0.2269 in the high-variance group, about 5.82 times larger.Source: 2) Comparison of Variable Admittance… · p. 13, Table V · p. 13
  • Under higher trajectory variance during the lifting rise, regulated stiffness decreases; as variance falls during descent or near motion end, stiffness increases.Interaction force and its fluctuations also increase with external load; the authors note that imperfect gravity compensation may contribute.Source: 2) Comparison of Variable Admittance… · p. 14, Figure 12 · p. 14
  • The RMS interaction force is lower than at 0 kg, 2 kg, and 5 kg: 3.7655 ± 1.7689 N, 8.3359 ± 5.4335 N, and 13.5434 ± 7.6466 N, respectively.The corresponding values for are 4.4275 ± 1.6690 N, 9.8248 ± 5.0079 N, and 14.6139 ± 6.4320 N.Source: 2) Comparison of Variable Admittance… · p. 14
  • The proposed controller yields lower activation than for all measured muscles under all three loads, with significant reductions under 0 kg and more selective significance at 2 kg.At 5 kg, mean activation is lower for all three muscles; the reported reductions are consistent with the interaction-force analysis.Source: 2) Comparison of Variable Admittance… · p. 15, Table VI · p. 15

3) Comparison of Overall Assistance Performance:

During repetitive 5 kg box lifting and placing, exoskeleton assistance reduced energy expenditure relative to natural motion. The conclusion describes an uncertainty-aware framework coupling multimodal trajectory generation with variable admittance control.

3) Comparison of Overall Assistance Performance:

  • Seven subjects completed randomized natural-motion and assisted conditions for four 2-minute sessions, separated by 40-second rests, during upright-trunk box lifting and placing.Respiratory measurements were converted to body-mass-normalized energy expenditure using the Brockway equation.Source: 3) Comparison of Overall Assistance… · p. 15
  • The first session showed an initial metabolic rise that approached a quasi-steady state after about 2 minutes, so it was excluded from the session comparison.The first-session transient motivated focusing the session comparison on the last three sessions.Source: 3) Comparison of Overall Assistance… · p. 15
  • 4.50 ± 0.61 to 4.08 ± 0.44 W · kg−1: assistance reduced full-trial average energy expenditure by 9.0% (p = 0.0041).Total energy expenditure fell from 184.93 ± 25.68 kJ to 167.64 ± 18.34 kJ.Source: Figure 13 · p. 15, 3) Comparison of Overall Assistance… · p. 15
  • 9.9%, 7.0%, and 9.4%: energy expenditure was lower with assistance in each of the last three sessions, with all reductions statistically significant.The reported p-values were 0.0081, 0.0030, and 0.0272, respectively.Source: Figure 14 · p. 16, 3) Comparison of Overall Assistance… · p. 16

V. CONCLUSION

  • The framework couples multimodal trajectory generation with variable admittance control; the conclusion reports lower muscle activation than baselines and lower metabolic expenditure during repetitive box lifting.The authors identify validation on multi-joint exoskeletons in more complex real-world assistance conditions as future work.Source: §V · p. 16

APPENDIX A IMPLEMENTATION DETAILS

The appendix summarizes implementation settings and describes the scope of generator and controller comparisons.

APPENDIX A IMPLEMENTATION DETAILS

  • Table VII summarizes generator windows, flow-matching settings, network and training choices, sensor processing, admittance parameters, and real-time latency.These parameters are presented to facilitate reproducibility.Source: Appendix A · p. 16, Table VII · p. 16

GENERATOR COMPARISON

  • The generator comparison assesses muscle-activation reduction against across tasks, with detailed task–muscle statistics reported in Appendix Table VIII.The appendix table supports the significance analysis in Table IV.Source: GENERATOR COMPARISON · p. 16

APPENDIX C DETAILED STATISTICAL RESULTS OF CONTROLLER

  • The controller comparison assesses muscle-activation reduction against under different external loads, with detailed statistics reported in Appendix Table IX.The appendix table supports the significance analysis in Table VI.Source: Appendix C · p. 16
Loading 2610.31302v1…