Imitation Learning from Human Motion Alone Does Not Guarantee Biomechanically Plausible Gait Kinetics

arXiv:2603.12408 · cs.RO, cs.LG · Submitted 2026-03-12 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: We're looking at the paper: "Imitation Learning from Human Motion Alone Does Not Guarantee Biomechanically Plausible Gait Kinetics".

Paper summary: Rosa: So we're looking at a paper called "Imitation Learning from Human Motion Alone Does Not Guarantee Biomechanically Plausible Gait Kinetics." Basically, they're asking if you can just use motion data to learn joint moments without ever knowing the actual forces involved in foot contact.

Dev: That sounds like a fundamental question for anyone building robots or even designing better exoskeletons.

Rosa: Exactly. The main point they make is that while imitation learning from motion works well for tracking how the body moves, it doesn't automatically mean you’ve learned the right internal forces, which are what we need to understand how a person actually walks.

Taro: So you're saying motion alone isn't enough to get those realistic joint moments?

Dev: That’s right. They compared a motion-only imitation learning approach, called MOIL, against something they called KAIL, which adds rewards for matching ground reaction forces and the center of pressure.

Rosa: And what did they find when they put those kinetic rewards in?

Dev: Well, both MOIL and KAIL tracked the motion—the kinematics—pretty well. But the paper says MOIL had substantially larger errors in things like ground reaction forces, center of pressure trajectories, and joint moment estimates compared to what inverse dynamics calculations predicted.

Taro: So even though they looked similar on the surface, there was a big difference in how they handled the actual physical interaction with the ground.

Rosa: That's really important for us because if you’re building something like a wearable robot, you need those joint moments to make sure it moves safely and plausibly.

Dev: It highlights that without knowing about those foot-ground interactions, the imitation learning process might just learn some pattern of motion that looks right kinematically but isn't physically possible.

Taro: I mean if the learned joint moments are off, you can end up with a policy that looks smooth in simulation but fails completely when it tries to interact with a real environment where things aren't perfect.

Rosa: So the researchers set up this comparison using walking data from a non-disabled person at three different speeds, zero point nine meters per second, one point two meters per second, and one point five meters per second.

Dev: Right. And they used this progressive kinetic reward shaping approach to see exactly how much adding those ground reaction forces or center of pressure matched the reference inverse dynamics outcomes.

Rosa: It sounds like they were trying to isolate the impact of adding those kinetics-aware rewards, right?

Dev: That's what they did. They started with a baseline policy trained only on kinematics for seven hundred iterations and then continued training for another three hundred iterations under four different reward configurations: kinematics only, kinematics plus ground reaction force matching, kinematics plus center of pressure matching, and finally the combined version called KAIL.

Taro: So when they looked at the results from this study on "Imitation Learning from Human Motion Alone Does Not Guarantee Biomechanically Plausible Gait Kinetics," what did they actually show about those reward configurations?

Paper summary: Rosa: They found that KAIL produced smoother profiles that aligned more closely with the reference inverse dynamics outcomes, especially at the hip and the knee.

Dev: And they reported an average improvement of fifty-four percent in external kinetics performance for the KAIL condition compared to something else.

Taro: That sounds like a solid improvement when you’re dealing with things like GRFs and CoP profiles in walking data.

Rosa: It really shows that while motion imitation alone doesn't guarantee those realistic joint moments, incorporating information about the foot-ground interaction makes the learned gait much more plausible biomechanically.

Dev: So even though KAIL showed better results, the paper still points out that the generated joint moments still differed from inverse dynamics results, which means there's still a model mismatch somewhere in the setup.

Taro: I mean, they acknowledge that this isn't a complete solution yet and point out some limitations. For instance, they only looked at lower limb kinematics and simplified the foot geometry in their simulation to focus on lower limb joint biomechanics consistent with how inverse dynamics is often done.

Rosa: That's a fair point. The model setup matters when you’re trying to generalize this outside of a controlled lab environment, right?

Dev: Right. And they also mentioned that the simplified foot model without a metatarsophalangeal joint might have constrained accurate reproduction of ground reaction forces and center of pressure profiles.

Taro: So the limitation isn't just in the learning, but in how much they can trust these results when they try to apply them to real-world scenarios where those contact details are more complex.

Rosa: And they also noted that the non-uniqueness problem in inverse dynamics means that similar joint motion can happen from different external contact wrenches and joint moments, so accurate contact measurements are needed to properly constrain the solution space for those joint moments.

Dev: It puts a lot of emphasis on needing more than just motion data to get the full picture of human gait biomechanics.

Taro: So what does this mean for us in terms of real-world application? If we want AI systems that can safely interact with humans, we need to move beyond just tracking body movement and start incorporating explicit constraints about how those feet are actually interacting with the ground.

Rosa: It means that for any system trying to mimic human walking or control a robot, relying only on motion data is risky because you might miss the critical kinetic details that make it feel and act like a real person.

Dev: Exactly. The work by Liu et al., in "Imitation Learning from Human Motion Alone Does Not Guarantee Biomechanically Plausible Gait Kinetics," shows us that while imitation learning is useful for generating stable, continuous movements in forward dynamics simulations, the underlying assumption that it can produce a physiologically plausible model of human gait might not be completely valid.

Conclusion: Rosa: So we're wrapping up this look at the paper "Imitation Learning from Human Motion Alone Does Not Guarantee Biomechanically Plausible Gait Kinetics." Basically, they found that just using motion data to learn how a person walks doesn't automatically mean you’ve learned the right internal forces or joint moments.

Dev: Right. The authors are arguing that for something like gait biomechanics, you need more than just watching someone move; you need to know what the ground is actually doing back on them.

Taro: And they tested this by comparing a simple imitation learning approach, MOIL, against one where they added rewards for matching ground reaction forces and center of pressure trajectories.

Rosa: The main finding is that while both methods could track the motion—the kinematics—well, the MOIL method had way bigger errors when it came to those actual forces and joint moments compared to what physics calculations predicted.

Dev: So, even though they looked similar on paper for the movement itself, the learned internal mechanics were just not matching up with reality.

Taro: That suggests that if you're building a robot or designing a prosthetic, relying only on motion data could lead to something that looks smooth in simulation but fails when it actually has to interact with the world.

Rosa: It implies that for any system trying to mimic human walking, incorporating explicit information about foot-ground interaction is absolutely necessary for it to be truly plausible.

Dev: The authors are pushing back on the idea that motion imitation alone is enough, stressing that without those kinetic details, you can't get a valid understanding of human gait.

Taro: That means the next step in making these systems useful involves getting better at measuring and using those contact forces accurately to constrain what the AI learns.

Rosa: It sets up a big question for us about where imitation learning is actually safe and effective when applied to complex, real-world physics.

University of North Carolina - Chapel Hill · North Carolina State University · Arizona State University

cs.RO, cs.LG

Submitted: 2026-03-12

Updated: 2026-11-18

Comments: 8 pages, 7 figures

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 66/100

The gist: Imitation learning (IL) via motion data is increasingly used in robotics and character animation, yet its ability to be applied to human biomechanics to estimate physiologically plausible joint…

Key concepts

Imitation Learning (IL)
A machine learning method where an agent learns to replicate expert behavior by observing demonstrations. It uses motion data to train a policy, aiming to mimic the actions of a human or expert system in simulations.
Joint Moments
These are internal forces calculated at each joint in the body during movement. They represent the torques required for muscles to produce specific joint motions, and accurately estimating them is crucial for understanding how human gait works physically.
Ground Reaction Forces (GRFs)
GRFs are the forces exerted by the ground on a body during activities like walking. These forces are vital because they directly influence the internal joint moments, making them necessary inputs for simulating realistic human movement kinetics.

Terminology

Summary

Imitation learning (IL) via motion data is increasingly used in robotics and character animation, yet its ability to be applied to human biomechanics to estimate physiologically plausible joint moments without explicit kinetic information remains unclear While both MOIL and KAIL achieved comparable kinematic tracking accuracy, MOIL exhibited substantially larger errors in GRFs, CoP trajectories, and joint moment estimates relative to inverse dynamics calculations as the references These findings highlight a fundamental limitation of MOIL approaches, as their failure to accurately estimate human-like gait kinetics may lead to erroneous interpretations of gait biomechanics and adversely affect downstream applications such as wearable robot design and the diagnosis of pathological gait

Study Objective

The objective of this study is to investigate whether MOIL is sufficient to learn policies that produce biomechanically plausible joint moments in walking The authors argue that additional kinetics-related measurements, such as foot contact forces, remain essential for estimating human-like gait kinetics To address this question, they extended MOIL to kinetics-aware IL (KAIL), which adds rewards for matching GRFs and CoP during normal walking

Learning Framework

The imitation learning (IL) problem is formulated as a finite-horizon Markov decision process (MDP) over a multi-segment dynamics model The MDP definition specifies the state vector st = [q−t, q˙r t, ϕt] constructed from the simulated environment The action at t is defined as a concatenation of equilibrium joint angles ψeq t and residual wrench ξt

Reward Design

The reward is decomposed into kinematics-related terms, aggregated as Rk, and additional kinetics-related terms that encourage agreement with the expert’s kinetic measurements The kinematics-related reward Rk is defined as the weighted sum of multiple kinematic imitation objectives using a negative-exponential reward Kinetics-related rewards include Ground Reaction Force (GRF) and Center of Pressure (CoP) rewards The GRF reward is defined as Rgrf = exp − α grf ∆f grf 2, where ∆f grf stacks the difference of sagittal plane GRFs

Experimental Results

The KAIL condition produced smoother profiles that more closely aligned reference inverse dynamics outcomes, notably at the hip and the knee The KAIL condition achieved an average improvement of 54% in external-kinetics performance and showed close agreement with inverse-dynamics joint moment estimates However, the generated joint moments still differed from the inverse-dynamics results, indicating model mismatch

Conclusion

The results imply that MOIL does not guarantee the production of human-like kinetics in gait simulations, whereas KAIL yields more biomechanically plausible internal and external kinetics The authors advocate that IL still requires critical information about foot-ground interactions to explain valid human gait biomechanics Consistent with the identifiability problem in inverse dynamics, accurate contact measurements are needed to constrain the solution space of joint moments further The paper concludes that IL-learned control policies remain valuable for their ability to generate stable and continuous human movements in forward dynamics simulations The underlying assumption that MOIL can produce a physiologically plausible model of human gait may not be completely valid

Limitations

A few limitations should be acknowledged, including the fact that only lower limb kinematics were considered in this study The model was simplified to focus on lower limb joint biomechanics, consistent with common inverse dynamics approaches in gait biomechanics This study was limited in scope to demonstrating the potential misuse of IL for accurate gait kinetics estimation Drawing population-level inferences about joint kinetics using an IL framework would require validation across multiple participants which will be addressed in future work The simplified foot geometry in the MuJoCo box-shaped foot model without a metatarsophalangeal joint likely constrained accurate reproduction of GRFs and CoP profiles The model used an impedance controller with preset stiffness and damping coefficients, which may yield jerkier joint moment profiles with higher frequency content The resulting joint moments may retain mechanical features that differ from biological kinetics even when external kinetics are encouraged through reward terms The nonconvex nature of reinforcement learning optimization can lead to residual kinetics errors and tradeoff effects Trade-offs between kinetic objectives were also observed, such as CoP progression in Fig. 6(d), as GRF and CoP impose distinct and sometimes competing physical reward objectives The simplified footground interaction model likely constrained accurate reproduction of GRFs and CoP profiles The non-uniqueness in inverse dynamics means similar joint motion can arise from different external contact wrenches and joint moments The paper is 2603. p. 6

Where the paper says it

6(d)), as GRF and CoP impose distinct and sometimes competing physical reward objectives; thus, joint optimization does not guarantee improvement in every metric. These discrepancies may first be attributed to the simplified footground interaction model.

12408.

--- Page 1 ---

Imitation learning (IL) via motion data is increasingly used in robotics and character animation, yet its ability to be applied to human biomechanics to estimate physiologically plausible joint moments without explicit kinetic information remains unclear

Introduction

Imitation learning (IL), also known as learning from demonstrations, is a machine learning paradigm in which an agent learns a policy by observing demonstrations from an expert It offers a promising approach to replicate demonstrated behaviors in forward dynamics simulations RL-based IL has significantly advanced applications across various domains, including robotic manipulation [4, 5, 6], physics-based animation [7, 8], autonomous driving [9, 10], humanoid control [11, 12] and assistive systems [13, 14]

Human Dynamics Model

The human body is modeled as a floating-base, multisegment system The system configuration is parameterized by the generalized coordinates q = [p, q, ψ]⊤ Forward dynamics simulations are implemented in MuJoCo [29], following the equations of motion for a constrained floating-base system Ground contact is represented through a set of holonomic constraints with Jacobian Jc(q), and f grf stacks the contact wrenches at the feet

Imitation Learning via Progressive Kinetic Reward Shaping Design

The authors employed a progressive, two-stage, kinetic reward shaping approach to evaluate the contribution of the additional reward components Specifically, all experiments shared a common baseline policy trained with kinematics reward terms for the first 700 iterations Starting from this identical checkpoint, training was continued for an additional 300 iterations under four reward configurations: (i) kinematics only (Baseline/MOIL), (ii) kinematics+GRF (GRF Match), (iii) kinematics+CoP (CoP Match), and (iv) kinematics+GRF+CoP (KAIL)

Model Evaluation and Analysis

The authors evaluated the kinematic accuracy, external kinetics fidelity, and internal kinetics plausibility of the four conditions using outcome signals such as joint angles, center-of-mass (CoM) velocities, GRFs, CoP trajectories, and joint moments Joint moments were normalized by body weight

Results

Across walking speeds, all conditions completed the full walking trials without falls and the aggregated kinematic reward Rk remained consistently within 0.82–0.85 regardless of the addition of kinetics-aware rewards The GRF Match condition reduced the total RMSE by 38%, while KAIL achieved an average improvement of 54% in external-kinetics performance KAIL produced smoother profiles that more closely aligned reference inverse dynamics outcomes, notably at the hip and the knee The KAIL condition achieved a reduction of 52% in CoP RMSE

Discussion

The results show that although MOIL reproduced human gait kinematics, it did not reliably produce human-like joint moments or foot-ground interactions Since joint moment calculation in walking is sensitive to foot contact force and location during the stance phase, when they added the measurements of GRF and CoP location into the reward function during training (i.e., KAIL), the policy yielded not only desired joint motion accurately but joint moments with reduced averaged RMSE by 40% at the hip, 34% at the knee, and 6% at the ankle in forward simulations of walking The results imply that MOIL does not guarantee the production of human-like kinetics in gait simulations KAIL yields more biomechanically plausible internal and external kinetics

References

[1] Brenna D Argall et al. “A survey of robot learning from demonstration”. p. 6

Where the paper says it

REFERENCES [1] Brenna D Argall et al. “A survey of robot learning from demonstration”.

In: Robotics and autonomous systems 57. p. 6
Where the paper says it

“A survey of robot learning from demonstration”. In: Robotics and autonomous systems 57.

5 (2009), pp. p. 6
Where the paper says it

In: Robotics and autonomous systems 57. 5 (2009), pp.

469–483

[2] Tengteng Zhang and Hongwei Mo. “Reinforcement learning for robot research: A comprehensive review and open issues”. p. 6

Where the paper says it

[2] Tengteng Zhang and Hongwei Mo. “Reinforcement learning for robot research: A comprehensive review and open issues”.

EN. In: International Journal of Advanced Robotic Systems 18. p. 6
Where the paper says it

“Reinforcement learning for robot research: A comprehensive review and open issues”. In: International Journal of Advanced Robotic Systems 18.

3 (May 2021). p. 6
Where the paper says it

In: International Journal of Advanced Robotic Systems 18. 3 (May 2021).

Publisher: SAGE Publications, p. p. 6
Where the paper says it

3 (May 2021). Publisher: SAGE Publications, p.

17298814211007305 ISSN: 1729-8806 DOI: 10. p. 6
Where the paper says it

17298814211007305. ISSN: 1729-8806.

1177/

[3] Maryam Zare et al. “A Survey of Imitation Learning: Algorithms, Recent Developments, and Challenges”. p. 6

Where the paper says it

[3] Maryam Zare et al. “A Survey of Imitation Learning: Algorithms, Recent Developments, and Challenges”.

In: IEEE Transactions on Cybernetics 54. p. 6
Where the paper says it

“A Survey of Imitation Learning: Algorithms, Recent Developments, and Challenges”. In: IEEE Transactions on Cybernetics 54.

12 (2024), pp. p. 6
Where the paper says it

In: IEEE Transactions on Cybernetics 54. 12 (2024), pp.

7173–7186 DOI: 10. p. 6
Where the paper says it

12 (2024), pp. 7173–7186.

1109 / TCYB. p. 6
Where the paper says it

7173–7186. 1109 / TCYB .

2024.

[4] Siheng Zhao et al. p. 6

Where the paper says it

1109 / TCYB . [4] Siheng Zhao et al.

ResMimic: From General Motion Tracking to Humanoid Whole-body LocoManipulation via Residual Learning. p. 6
Where the paper says it

[4] Siheng Zhao et al. ResMimic: From General Motion Tracking to Humanoid Whole-body LocoManipulation via Residual Learning.

2025. arXiv: 2510. p. 6
Where the paper says it

ResMimic: From General Motion Tracking to Humanoid Whole-body LocoManipulation via Residual Learning. arXiv: 2510.

05070 [cs].

[5] Yuke Zhu et al. p. 6

Where the paper says it

arXiv: 2510. [5] Yuke Zhu et al.

Reinforcement and Imitation Learning for Diverse Visuomotor Skills. p. 6
Where the paper says it

[5] Yuke Zhu et al. Reinforcement and Imitation Learning for Diverse Visuomotor Skills.

2018. arXiv: 1802. p. 6
Where the paper says it

Reinforcement and Imitation Learning for Diverse Visuomotor Skills. arXiv: 1802.

[6] Liyiming Ke et al. p. 6

Where the paper says it

arXiv: 1802. [6] Liyiming Ke et al.

“Grasping with Chopsticks: Combating Covariate Shift in Model-free Imitation Learning for Fine Manipulation”. p. 6
Where the paper says it

[6] Liyiming Ke et al. “Grasping with Chopsticks: Combating Covariate Shift in Model-free Imitation Learning for Fine Manipulation”.

In: 2021 IEEE International Conference on Robotics and Automation (ICRA). p. 6
Where the paper says it

“Grasping with Chopsticks: Combating Covariate Shift in Model-free Imitation Learning for Fine Manipulation”. In: 2021 IEEE International Conference on Robotics and Automation (ICRA).

Xi’an, China: IEEE Press, 2021, pp. p. 6
Where the paper says it

In: 2021 IEEE International Conference on Robotics and Automation (ICRA). Xi’an, China: IEEE Press, 2021, pp.

6185–6191 DOI: 10. p. 6
Where the paper says it

Xi’an, China: IEEE Press, 2021, pp. 6185– 6191.

1109 / ICRA48506. p. 6
Where the paper says it

6185– 6191. 1109 / ICRA48506 .

[7] Xue Bin Peng et al. p. 6

Where the paper says it

1109 / ICRA48506 . [7] Xue Bin Peng et al.

“DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills”. p. 6
Where the paper says it

[7] Xue Bin Peng et al. “DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills”.

In: CoRR abs/1804. p. 6
Where the paper says it

“DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills”. In: CoRR abs/1804.

02717 (2018). p. 6
Where the paper says it

In: CoRR abs/1804. 02717 (2018).

arXiv: 1804. p. 6
Where the paper says it

02717 (2018). arXiv: 1804.

02717.

[8] Ye Yuan and Kris Kitani. p. 6

Where the paper says it

arXiv: 1804. [8] Ye Yuan and Kris Kitani.

“Residual force control for agile human behavior imitation and extended motion synthesis”. p. 6
Where the paper says it

[8] Ye Yuan and Kris Kitani. “Residual force control for agile human behavior imitation and extended motion synthesis”.

In: Advances in Neural Information Processing Systems 33 (2020), pp. p. 6
Where the paper says it

“Residual force control for agile human behavior imitation and extended motion synthesis”. In: Advances in Neural Information Processing Systems 33 (2020), pp.

21763–21774 DOI: 10. p. 6
Where the paper says it

In: Advances in Neural Information Processing Systems 33 (2020), pp. 21763–21774.

1186/s

[9] Jianyu Chen, Bodi Yuan, and Masayoshi Tomizuka. “Deep Imitation Learning for Autonomous Driving in Generic Urban Scenarios with Enhanced Safety”. p. 6

Where the paper says it

[9] Jianyu Chen, Bodi Yuan, and Masayoshi Tomizuka. “Deep Imitation Learning for Autonomous Driving in Generic Urban Scenarios with Enhanced Safety”.

In: 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). p. 6
Where the paper says it

“Deep Imitation Learning for Autonomous Driving in Generic Urban Scenarios with Enhanced Safety”. In: 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS).

Macau, China: IEEE Press, 2019, pp. p. 6
Where the paper says it

In: 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). Macau, China: IEEE Press, 2019, pp.

2884–2890 DOI: 10. p. 6
Where the paper says it

Macau, China: IEEE Press, 2019, pp. 2884–2890.

1109/IROS40897. p. 6
Where the paper says it

2884–2890. 1109/ IROS40897.

[10] Jinning Li et al. p. 6

Where the paper says it

1109/ IROS40897. [10] Jinning Li et al.

“Guided Online Distillation: Promoting Safe Reinforcement Learning by Offline Demonstration”. p. 6
Where the paper says it

[10] Jinning Li et al. “Guided Online Distillation: Promoting Safe Reinforcement Learning by Offline Demonstration”.

In: 2024 IEEE International Conference on Robotics and Automation (ICRA). p. 6
Where the paper says it

“Guided Online Distillation: Promoting Safe Reinforcement Learning by Offline Demonstration”. In: 2024 IEEE International Conference on Robotics and Automation (ICRA).

2024, pp. 7447–7454 DOI: 10. p. 6
Where the paper says it

In: 2024 IEEE International Conference on Robotics and Automation (ICRA). 7447– 7454.

1109 / ICRA57147. p. 6
Where the paper says it

7447– 7454. 1109 / ICRA57147 .

[11] Lujie Yang et al. p. 6

Where the paper says it

1109 / ICRA57147 . [11] Lujie Yang et al.

OmniRetarget: InteractionPreserving Data Generation for Humanoid WholeBody Loco-Manipulation and Scene Interaction. p. 6
Where the paper says it

[11] Lujie Yang et al. OmniRetarget: InteractionPreserving Data Generation for Humanoid WholeBody Loco-Manipulation and Scene Interaction.

arXiv:2509. p. 6
Where the paper says it

OmniRetarget: InteractionPreserving Data Generation for Humanoid WholeBody Loco-Manipulation and Scene Interaction. arXiv:2509.

26633 [cs]. p. 6
Where the paper says it

arXiv:2509. 26633 [cs].

Oct. 2025 DOI: 10. p. 7
Where the paper says it

In: arXiv preprint arXiv:2501. 02116 (2025). p. 7

Where the paper says it

In: arXiv preprint arXiv:2501. 02116 (2025).

48550/arXiv. p. 6
Where the paper says it

26633 [cs]. 48550/ arXiv.

[12] Zhaoyuan Gu et al. p. 7

Where the paper says it

[12] Zhaoyuan Gu et al. “Humanoid locomotion and manipulation: Current progress and challenges in control, planning, and learning”.

“Humanoid locomotion and manipulation: Current progress and challenges in control, planning, and learning”. p. 7
Where the paper says it

[12] Zhaoyuan Gu et al. “Humanoid locomotion and manipulation: Current progress and challenges in control, planning, and learning”.

In: arXiv preprint arXiv:2501. p. 7
Where the paper says it

“Humanoid locomotion and manipulation: Current progress and challenges in control, planning, and learning”. In: arXiv preprint arXiv:2501.

02116 (2025).

--- Page 3 ---

Fig. 3. Detailed biomechanics imitation-learning pipeline corresponding to the dashed box in Fig. p. 2

Where the paper says it

Detailed biomechanics imitation-learning pipeline corresponding to the dashed box in Fig. The proximal policy optimization (PPO) agent learns to imitate expert gait from rewards defined in Sec.

1(b)

Problem Formulation

The imitation learning (IL) problem is formulated as a finite-horizon Markov decision process (MDP) over a multi-segment dynamics model The MDP definition specifies the state vector st = [q−t, q˙r t, ϕt] constructed from the simulated environment The action at t is defined as a concatenation of equilibrium joint angles ψeq t and residual wrench ξt

Training Setup

Proximal Policy Optimization (PPO) was used to train the policy in the aforementioned environment The policy was modeled as a Gaussian stochastic actor with fixed action standard deviation exp(−2.3), implemented as a two-layer ReLU network with 512 and 256 hidden units Training employed PPO with clipping parameterϵc = 0. p. 2

Where the paper says it

The policy was modeled as a Gaussian stochastic actor with fixed action standard deviation exp(−2. 3), implemented as a two-layer ReLU network with 512 and 256 hidden units.

2 and generalized advantage estimation with discount factor γ = 0. p. 2
Where the paper says it

Training employed PPO with clipping parameter ϵc = 0. 2 and generalized advantage estimation with discount factor γ = 0.

95 and generalized advantage estimation mixing parameter τ = 0. p. 2
Where the paper says it

2 and generalized advantage estimation with discount factor γ = 0. 95 and generalized advantage estimation mixing parameter τ = 0.

95

--- Page 4 ---

Fig. 4. Progressive Kinetic Reward Shaping scheme and corresponding reward curves

Progressive Kinetic Reward Shaping Scheme

The progressive kinetic reward shaping approach was conducted to evaluate the contribution of each kinetics-aware reward term All experiments shared a common baseline policy trained with kinematics reward terms for the first 700 iterations Starting from this identical checkpoint, training was continued for an additional 300 iterations under four reward configurations: (i) kinematics only (Baseline/MOIL), (ii) kinematics+GRF (GRF Match), (iii) kinematics+CoP (CoP Match), and (iv) kinematics+GRF+CoP (KAIL)

Internal Kinetics Plausibility

Qualitatively, MOIL yielded irregular estimates at key gait events, particularly at heel-strike and toeoff, across all speeds In contrast, KAIL produced smoother profiles that more closely aligned reference inverse dynamics outcomes, notably at the hip and the knee Joint moment accuracy was evaluated for the MOIL and KAIL conditions at each walking speed using RMSE and Pearson’s R relative to the inverse dynamics reference moments

--- Page 5 ---

Fig. 5. Joint angle trajectories and tracking accuracy comparison

Kinematics Fidelity

The average joint angle RMSE across walking speeds was 3.66° for MOIL, 4. p. 0

Where the paper says it

In this study, we examined whether motion imitation alone can estimate reasonable biological joint moments in human walking. We compared motion-only IL (MOIL) against a kinetics-aware IL (KAIL) framework, which incorporates additional ground reaction forces (GRFs) or center of pressure (CoP) in imitation rewards.

40° for GRF Match, 4. p. 3
Where the paper says it

Specifically, all experiments shared a common baseline policy trained with kinematics reward terms for the first 700 iterations. Starting from this identical checkpoint, training was continued for an additional 300 iterations under four reward configurations: (i) kinematics only (Baseline/MOIL), (ii) kinematics+GRF (GRF Match), (iii) kinematics+CoP (CoP Match), and (iv) kinematics+GRF+CoP (KAIL).

33° for CoP Match, and 4. p. 3
Where the paper says it

Specifically, all experiments shared a common baseline policy trained with kinematics reward terms for the first 700 iterations. Starting from this identical checkpoint, training was continued for an additional 300 iterations under four reward configurations: (i) kinematics only (Baseline/MOIL), (ii) kinematics+GRF (GRF Match), (iii) kinematics+CoP (CoP Match), and (iv) kinematics+GRF+CoP (KAIL).

02° for KAIL The instantaneous CoM sagittal plane velocity RMSEs across speeds and kinetic reward shaping conditions were shown in Fig. p. 4
Where the paper says it

02° for KAIL. The instantaneous CoM sagittal plane velocity RMSEs across speeds and kinetic reward shaping conditions are shown in Fig.

5(d) The GRF Match condition had mean average RMSE of 0. p. 4
Where the paper says it

Averaged across three speeds, The GRF Match condition reduced the total RMSE by 38%, while the CoP Match condition yielded an improvement of 41%. KAIL achieved an average improvement of 54%.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Incorporate kinetics-aware imitation learning (KAIL) frameworks into reinforcement learning agents to estimate physiologically plausible joint moments by incorporating additional ground reaction forces (GRFs) or center of pressure (CoP) in imitation rewards. This allows the agent to learn policies that produce kinetics more closely aligned with biomechanical values.

  2. Implement a progressive kinetic reward shaping approach during training, where the reward function is decomposed into kinematic terms and kinetics-related terms, such as Ground reaction force (GRF) reward and Center of Pressure (CoP) reward, to evaluate the contribution of each term to policy refinement. This structured training protocol isolates the incremental contribution of kinetic objectives.

  3. The improved AI system can perform more accurate biomechanical gait analysis by generating joint moments with reduced average root mean square error, specifically showing a reduction in RMSE by 40% at the hip, 34% at the knee, and 6% at the ankle compared to motion-only imitation learning.

  4. The system can achieve better external kinetics fidelity by improving GRF estimates and CoP trajectories through explicit kinetic rewards, as demonstrated by KAIL achieved an average improvement of 54% in GRF RMSE across three speeds.

  5. The improved AI system can generate joint moment profiles that are more physically plausible, showing smoother profiles that more closely aligned reference inverse dynamics outcomes in KAIL conditions, addressing the limitation where MOIL yielded irregular estimates at key gait events.

Abstract

Imitation learning (IL) via motion data is increasingly used in robotics and character animation, yet its ability to be applied to human biomechanics to estimate physiologically plausible joint moments without explicit kinetic information remains unclear. In this study, we examined whether motion imitation alone can estimate reasonable biological joint moments in human walking. We compared motion-only IL (MOIL) against a kinetics-aware IL (KAIL) framework, which incorporates additional ground reaction forces (GRFs) and center of pressure (CoP) in imitation rewards. A progressive kinetic reward shaping approach was used to examine the contribution of each kinetic term. Experiments were conducted using walking data from two non-disabled participants at three speeds (0.9, 1.2, and 1.5 m/s). While both MOIL and KAIL achieved comparable kinematic tracking accuracy, KAIL produced kinetics more closely aligned with biomechanical values across participants and walking speeds, reducing the root mean square errors (RMSEs) of vertical GRF and CoP by 40% and 48%, and of hip, knee, and ankle moments by 19%, 18%, and 35%, respectively, relative to MOIL. These findings highlight a fundamental limitation of MOIL approaches, as their failure to accurately estimate human-like gait kinetics may lead to erroneous interpretations of gait biomechanics and adversely affect downstream applications such as wearable robot design and the diagnosis of pathological gait.

Sources

Related papers