Recovering the forces behind a motion is ill-posed: different controllers can reproduce the same movement while generating different contact forces. Physics-based motion imitation resolves this ambiguity through simulation, but relies on fixed-gain proportional-derivative (PD) control. We show that fixed gains impose a trade-off between motion and force tracking, as contact and swing place conflicting demands on joint stiffness and damping. We introduce ImpedanceMimic, a reinforcement-learning policy that modulates per-joint stiffness, damping, and position setpoints from motion and force measurements, but requires only motion at inference. Across 129 participants and six activities, ImpedanceMimic outperforms matched fixed-gain PD under every integration scheme, direct-torque, muscle-actuated and existing imitation controllers in kinematic and force fidelity, approaching a dedicated motion-to-force regressor. By expressing gains as multiples of inertia-scaled priors, a single policy controls subject-specific biomechanical models without explicit anthropometric input and transfers zero-shot to unseen humans and a humanoid robot. The learned impedance forms a low-dimensional representation of motor state, shared across bodies, that generalizes to withheld activities: stiffness scales with load, and damping anticipates unloading.
Movement is governed by forces, which in locomotion arise from foot–ground contact and determine a sprinter’s acceleration, a patient’s loading asymmetry and a humanoid robot’s balance. Cameras have made motion datasets abundant [22, 34, 65], but force data remains scarce. Force plates confine recordings to the laboratory, and wearable insoles resolve only the normal component and drift over time [23]. Recovering forces from motion alone would extend movement analysis into natural, unconstrained settings.
Recent corpora synchronize motion capture with force-plate measurements [61], extending beyond steady gait to tasks like cutting, lunging and jumping [46]. Learning-based methods use such data to map kinematics to ground reaction forces (GRF), either through direct regression [39] or generative modeling [55]. Their predictions can violate physics, for example by loading the ground during flight or placing the center of pressure (CoP) outside the foot, since a loss term at best encourages physical consistency. The learned mapping is also opaque, going from motion to force directly without modeling the actuation and body dynamics that produce it, and therefore cannot drive a simulated body or robot.
Physics-based motion imitation addresses both limitations. Tracking a recorded motion in a physics engine yields contact forces that are dynamically consistent by construction [41, 31], and the controller producing these forces is explicitly defined. However, dynamic consistency does not imply agreement with measured forces. Most imitation policies output joint-position targets, which a proportional–derivative (PD) controller with hand-tuned gains converts into joint torques to drive the simulated body [40]. Different gains can track the same motion, yet produce different contact forces. Recent work incorporates measured force as an observation [20] or a reward [29], but fixes the gains and reports accuracy either with force sensors available at inference or on a single participant. Alternative actuation models, including learned gains [63], muscle models [27] and direct torque control [10], have not been systematically evaluated against measured forces.
Here, we present ImpedanceMimic, a reinforcement-learning imitation policy that outputs per-joint stiffness, damping and position setpoints, supervised by reference kinematics, force-plate GRF and CoP. On 129 participants, nine datasets and six activities, ImpedanceMimic outperforms fixed-PD, direct-torque, muscle and existing imitation controllers on kinematic and force accuracy. To our knowledge, no prior controller reproduces measured per-foot contact forces across subjects and activities, including unseen ones, without force at inference. We contribute:
Action-space bounds on force recovery. We derive a lower bound on stiffness, set by the torque the reference force demands, and an upper bound on damping, set by discrete-time stability, which oppose each other between contact and swing. Empirically, a fixed-gain policy forfeits either motion or force tracking, whereas learned impedance satisfies both, extending the action-space analysis of Peng & Van De Panne [40] to force.
A morphology-agnostic controller. Expressing impedance as a scaled inertia prior, our policy, trained on 107 subject-specific models without anthropometric input, transfers zero-shot to 22 subjects and a simulated humanoid robot (G1). To our knowledge, this is the first physics-based imitator to transfer across unseen biomechanical and robotic embodiments.
Learned impedance encodes motor state. The learned gains form a low-dimensional, interpretable representation in which stiffness scales with load and damping anticipates unloading. It generalizes to withheld activities better than fixed action spaces and identifies the activity from a single cycle of gains. This representation carries over to unseen subjects.
An open benchmark. We release code, checkpoints and splits for a benchmark spanning 129 participants and six activities, with ablations, fixed-PD, direct-torque and muscle action spaces, and the existing methods PHC (Luo et al., [31]), SimPoE (Yuan et al., [63]) and GaitDynamics (Tan et al., [55]).
Related Work
Physics-based motion imitation. Executing a recorded motion in a simulator makes the resulting forces dynamically consistent. Reinforcement learning has made this routine at scale, from single-clip tracking (Peng et al., [41]) through adversarial objectives (Peng et al., [42]) to controllers covering large motion corpora (Luo et al., [31]; 2024; Tessler et al., [56]) and kinetic labeling (Liu et al., [28]). These methods pursue a kinematic objective because they originate in character animation, where simulation corrects generated motion (Yuan et al., [64]) rather than recovering the forces behind it. Comparing action spaces on motion-tracking tasks, Peng & Van De Panne [40] found fixed-gain PD control superior to torque, velocity and muscle parameterizations, and most later work adopted it. In biomechanics, trajectory optimization over musculoskeletal models (Dembia et al., [11]) estimates individual muscle forces, but each trial requires its own slow optimization. Recent work extends RL imitation to anatomically grounded models with torque or muscle actuation (Cotton, [10]; Simos et al., [49]; Li et al., [27]; Choi et al., [8]), with limited validation of the resulting forces against measured ones. Hori et al. [20] provide measured force as an observation and therefore need inertial and insole sensors at inference. Liu et al. [29] add GRF and CoP to the reward, but validate on a single walking participant and rely on an external pelvis wrench (Yuan & Kitani, [62]). ImpedanceMimic is trained on measured per-foot force and CoP from 129 participants and six activities, and needs neither sensors nor an external wrench at inference.
Learned impedance control. Letting a policy set its own impedance originates in robotics, where variable stiffness and damping were learned to make contact-rich manipulation and legged hopping robust to contact uncertainty [5, 35, 2], the contact force being a quantity to minimize rather than to reproduce. The idea has since entered physics-based character control and motion capture. SimPoE [63] scales nominal gains at every PD step and adds a residual root wrench to the action, ReGAIL [3] puts servo gains and damping in the action so that a controller can alternate tension and relaxation, and Neural PhysCap [48] and OSDCap [24] regress the gains from video, the latter together with an external force inside a neural Kalman filter. In all four, the gains serve motion imitation or pose estimation, without being tied to the controlled body or evaluated against measured force. Le et al. (2026) come closest, predicting the gains of a center-of-mass PD controller against force plates, though they distribute the force over body vertices instead of controlling joints. No prior work has learned the impedance of an articulated human model against measured contact forces. ImpedanceMimic learns joint impedance against measured force, with gains expressed as multiples of inertia-scaled priors that tie them to the body. We first motivate our approach on a single joint, which we then scale to human motion imitation.
Method
Ill-posed contact forces. We model the body as a floating-base rigid-body system with generalized coordinates q∈Rn, n=31, six of which describe the pelvis and m=15 of which are actuated joints. Its dynamics obey the Newton–Euler equations of motion,
M(q)q¨+h(q,q˙)=S⊤τ+Jc(q)⊤λ,(1)
with mass matrix M, bias forces h (Coriolis, centrifugal, gravitational), joint torques τ mapped by the selection matrix S, and contact forces λ through their Jacobian Jc (dimensions in App. § 9.2.2). Each GRF is the resultant of its contact forces applied at its CoP. A recorded motion q∗(t) fixes the left-hand side of Eq. (1), and its torque-free pelvis rows fix the net contact wrench to the rate of change of whole-body momentum, but not how that wrench is realized: in double support, its split between the feet and each CoP remains undetermined. Simulation resolves this through a contact model and a joint-torque trajectory τ(t), so the accuracy of the recovered forces is that of τ(t).
Action-space Impedance Bounds
A policy πθ(at∣ot) acts every Δt=10ms while the simulator steps every h=2ms. The action space determines how the joint torque τ^ in Eq. (1) evolves over the N=5 substeps: a direct-torque policy holds τ^=τscaleat fixed, so it cannot react between queries. Peng & Van De Panne (2017) wrap it in a proportional–derivative controller recomputed every substep, to smooth learning:
τ^=kp∘(qt+1∗+αδq−q)+kd∘(q˙t+1∗−q˙),(2)
clipped to ±τlim, where δq∈[−1,1]m are per-joint position offsets scaled by α, and ∘ the element-wise product. We extend this controller by emitting per-joint gain residuals Δkp,Δkd∈[−1,1]m,
kp=k0p∘exp(spΔkp),kd=k0d∘exp(sdΔkd),(3)
about base gains k0p,k0d, with log-scales sp,sd>0 setting the gain range. The exponential keeps gains strictly positive, as stability at the joint requires. We define two PD action spaces: Fixed-PD (a=δq, gains held at (k0p,k0d)), and Learned-PD (a=(Δkp,Δkd,δq)). As observed empirically by Bogdanovic et al. (2020), in contact the environment constrains the joint, so kd has little effect and kp sets the force the PD can produce. Out of contact the joint drives only the light distal segment, so kd governs the closed-loop dynamics. We formalize this as two bounds.
Proposition 1 (Force authority, Bound 1). At the reference state(q,q˙)=(q∗,q˙∗) the velocity-error term of Eq. (2) vanishes and the displacement is O(Δt)-small, so the PD torque reduces to τ≈αkp∘δq: the offset δq is the policy’s sole lever for torque production at the reference pose, giving an achievable interval of half-width αkp. Producing the reference contact force therefore requires the inverse-dynamics torque τ∗ of Eq. (1), which is attainable at the reference pose only if
kp≥αmaxt∣τ∗(t)∣(4)
Below this bound, the lever αkp is too small to supply the torque required by the reference force. The offset δq saturates, leaving the tracking term kp∘(q∗−q) as the only source of torque. A fixed-gain policy below the bound must therefore choose between forfeiting kinematic or force tracking.
Proposition 2 (Feedback stability, Bound 2). In continuous time any positive kd is stabilizing. In simulation the PD of Eq. (2) is applied as an explicit actuator force at every physics step h[40]. On a free joint of inertia I, discretizing the error dynamics yields a 2×2 linear map whose eigenvalues must lie inside the unit circle. The Jury criterion gives (App. A.2.1)
0<kd<h2Iand0<kp<h24I−h2kd.(5)
The damping condition is the more restrictive case, and depends on I, the changing inertia that the joint drives at a given moment. During contact the joint acts on the kinematic chain of the whole body above it, whereas in swing it drives only the distal segment, whose inertia is orders of magnitude smaller. A damping gain that is stable in contact can therefore be unstable in flight, where the feedback amplifies the error between policy steps and the torque alternates at ±τlim. This chatter is the classical limit on the impedance a sampled-data controller can render [9]. Under the explicit PD of Eq. (2), each joint thus inherits a damping ceiling set by its distal inertia. Stable PD avoids it by evaluating the controller at the next state [51], and implicit damping by moving the damping term into the integrator [57].
Remark 1 (Phase-dependent feasibility). A stiff joint, as Bound 1 demands in contact, needs higher damping against oscillation, whereas Bound 2 needs a lightly damped joint in swing. Fixed gains must therefore favor one phase over the other, whereas state-dependent gains can match each bound.
Validation. We test both bounds on the rod of Figure 2a, where they evaluate to kp≥32 Nm/rad and kd<20 Nm s/rad (App. A.2.1). Trained with identical base gains, reward and PPO [47], Fixed-PD forfeits force tracking when its stiffness is below Bound 1 (Figure 2b) and loses kinematic and force tracking for damping above Bound 2 (Figure 2c). Learned-PD instead reaches the same force accuracy by raising kp to Bound 1 in contact and holding kd under Bound 2 throughout (Figure 2d). Section 4.1 verifies these bounds on the human model (Figure 9, Table 16).
Figure 2. Action-space bounds on a single joint. (a) A hinged rod tracks a trajectory into a compliant floor, drawn from its swing apex (light) to contact (dark). (b) Force error of Learned-PD (red) and Fixed-PD (blue) versus base stiffness, with Bound 1 (B1) at 32 Nm/rad. (c) Jerk, angle and force error versus base damping, with Bound 2 (B2) at 20 Nm s/rad. (d) Both bounds, the Fixed-PD gains and the Learned-PD gains over one cycle, solid in contact and translucent in swing.
Figure 9. Action-space bounds on the joint-actuated human model. (a–c) Contact stiffness against the Bound-1 demand ∣τ∗∣/α. (d–f) Swing damping against Bound 2 on a log scale.
Joint
Bound 2 (×k0d)
walk B1 %
walk B2 %
run B1 %
run B2 %
stand B1 %
stand B2 %
lateral B1 %
lateral B2 %
flexion B1 %
flexion B2 %
jump B1 %
jump B2 %
hip
100.27
98 / 100
100 / 100
99 / 99
100 / 100
100 / 100
–
95 / 97
100 / 100
96 / 98
100 / 100
98 / 99
100 / 100
knee
16.43
98 / 100
100 / 100
97 / 72
100 / 100
100 / 100
–
97 / 89
100 / 100
97 / 100
100 / 100
98 / 99
100 / 100
ankle
0.77
96 / 93
100 / 0
92 / 53
100 / 0
100 / 100
–
79 / 76
99 / 0
100 / 100
100 / 0
94 / 94
96 / 0
Table 16. Action-space bounds on the joint-actuated human model. Shares (%) of contact frames satisfying Bound 1 (B1) and of swing frames satisfying Bound 2 (B2) on seen held-out clips. Each cell reads learned gains (mean over five seeds) / nominal gains. Standing has no swing phase.
ImpedanceMimic
Action space. The policy outputs at=(Δktp,Δktd,δqt)∈[−1,1]3m with log-scales sp=1.25 and sd=1.5. A first-order lag mimics activation dynamics by smoothing gain changes,
that would otherwise appear as force transients, and requires the policy to stiffen ahead of touchdown. Base gains scale with each subject i’s diagonal joint-space inertia Ijj(i) at the rest pose,
k0p=Ijjωn2,k0d=2ζ∘Ijj∘ωn,(7)
where the natural frequency ωn and damping ratio ζ are set per joint by nominal gains (Table 8), such that a multiple of the nominal results in the same closed-loop behavior across bodies.
References
[1]Åström, K. J., & Wittenmark, B. (2013). Computer-controlled systems: theory and design. In Courier Corporation.
[2]Miroslav Bogdanovic, Majid Khadiv, and Ludovic Righetti. Learning variable impedance control for contact sensitive tasks. IEEE Robotics and Automation Letters, 5(4):6129–6136, 2020.arxiv.org/abs/1907.07500
[3]Paul Boursin, Yannis Kedadry, Tony Chevalier, Victor Zordan, Paul Kry, Sophie Gregoire, and Marie-Paule Cani. Imitation in relative terms using ReGAIL: Making motion controllers agile and transferable. Computers & Graphics, 133:104457, 2025.DOI
[4]Sjoerd M Bruijn, Onno G Meijer, Peter J Beek, and Jaap H Van Dieen. The effects of arm swing on human gait stability. Journal of experimental biology, 213(23):3945–3952, 2010.DOI
[5]Buchli, J., Stulp, F., Theodorou, E., & Schaal, S. (2011). Learning variable impedance control. The International Journal of Robotics Research, 30(7), 820–833.
[6]Jonathan Camargo, Aditya Ramanathan, Will Flanagan, and Aaron Young. A comprehensive, open-source dataset of lower limb biomechanics in multiple conditions of stairs, ramps, and level-ground ambulation and transitions. Journal of biomechanics, 119:110320, 2021.DOI
[7]Josh Carter, Xi Chen, Dario Cazzola, Grant Trewartha, and Ezio Preatoni. Full body kinematics and ground reaction forces of fifty heterogeneous runners completing treadmill running at various speeds and gradients. University of Bath Research Data Archive, 2024. doi:10.15125/BATH-01341.DOI
[8]Itak Choi, Ilseung Park, Eni Halilaj, and Inseung Kang. Musculoskeletal motion imitation for learning personalized exoskeleton control policy in impaired gait. arXiv preprint arXiv:2604.09431, 2026.arXiv
[9]Colgate, J. E., & Brown, J. M. (1994). Factors affecting the Z-width of a haptic display. In Proceedings of the 1994 IEEE international conference on robotics and automation (pp. 3205–3210).
[10]R James Cotton. KinTwin: imitation learning with torque and muscle driven biomechanical models enables precise replication of able-bodied and impaired movement from markerless motion capture. arXiv preprint arXiv:2505.13436, 2025.arXiv
[11]Christopher L Dembia, Nicholas A Bianco, Antoine Falisse, Jennifer L Hicks, and Scott L Delp. OpenSim Moco: Musculoskeletal optimal control. PLoS computational biology, 16(12):e1008493, 2020.DOI
[12]Damiana A Dos Santos, Claudiane A Fukuchi, Reginaldo K Fukuchi, and Marcos Duarte. A data set with kinematic and ground reaction forces of human balance. PeerJ, 5:e3626, 2017.DOI
[13]Antoine Falisse, Gil Serrancolí, Christopher L Dembia, Joris Gillis, Ilse Jonkers, and Friedl De Groote. Rapid predictive simulations with complex musculoskeletal models suggest that diverse healthy and pathological human gaits can emerge from similar control strategies. Journal of The Royal Society Interface, 16(157): 20190402, 2019.DOI
[14]Claire T Farley and Octavio Gonzalez. Leg stiffness and stride frequency in human running. Journal of biomechanics, 29(2):181–186, 1996.pubmed.ncbi.nlm.nih.gov/8849811
[15]Benjamin J Fregly, Thor F Besier, David G Lloyd, Scott L Delp, Scott A Banks, Marcus G Pandy, and Darryl D D’lima. Grand challenge competition to predict in vivo knee loads. Journal of orthopaedic research, 30(4): 503–513, 2012.DOI
[16]DeepMind, G., & NVIDIA. (2025). MuJoCo Warp: GPU-optimized version of the MuJoCo physics simulator. <a href="https://doi.org/https://github.com/google-deepmind/mujoco_warp" rel="noopener noreferrer" target="_blank">https://doi.org/https://github.com/google-deepmind/mujoco_warp</a>
[17]Samuel R Hamner and Scott L Delp. Muscle contributions to fore-aft and vertical body mass center accelerations over a range of running speeds. Journal of biomechanics, 46(4):780–787, 2013.DOI
[18]Xingjian Han, Ben Senderling, Stanley To, Deepak Kumar, Emily Whiting, and Jun Saito. GroundLink: A dataset unifying human body movement and ground reaction dynamics. In SIGGRAPH Asia 2023 Conference Papers, pp. 1–10, 2023.DOI
[19]Neville Hogan. Adaptive control of mechanical impedance by coactivation of antagonist muscles. IEEE Transactions on automatic control, 29(8):681–690, 1984.DOI
[20]Ryosuke Hori, Jyun-Ting Song, Zhengyi Luo, Jinkun Cao, Soyong Shin, Hideo Saito, and Kris Kitani. Ground reaction inertial poser: Physics-based human motion capture from sparse IMUs and insole pressure sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026.arxiv.org/abs/2603.16233
[21]Aleksi Ikkala and Perttu Hämäläinen. Converting biomechanical models from OpenSim to MuJoCo. In Converging Clinical and Engineering Research on Neurorehabilitation IV: Proceedings of the 5th International Conference on Neurorehabilitation (ICNR2020), pp. 277–281. Springer, 2022.DOI
[22]Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6M: Large scale datasets and predictive methods for 3D human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325–1339, 2014.DOI
[23]Markus Koch, Lars-Kristian Lunde, Michael Ernst, Stein Knardahl, and Kaj Bo Veiersted. Validity and reliability of pressure-measurement insoles for vertical ground reaction force assessment in field situations. Applied Ergonomics, 53:44–51, 2016.DOI
[24]Cuong Le, Viktor Johansson, Manon Kok, and Bastian Wandt. Optimal-state dynamics estimation for physics-based human motion capture from videos. Advances in Neural Information Processing Systems, 37:43609–43631, 2024.arxiv.org/abs/2410.07795
[26]Hyunglae Lee and Neville Hogan. Time-varying angle mechanical impedance during human locomotion. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 23(5):755–764, 2015.DOI
[27]Chengkun Li, Cheryl Wang, Bianca Ziliotto, Merkourios Simos, Jozsef Kovecses, Guillaume Durandau, and Alexander Mathis. Towards embodied AI with MuscleMimic: Unlocking full-body musculoskeletal motor learning at scale. arXiv preprint arXiv:2603.25544, 2026.arXiv
[28]Xinpeng Liu, Junxuan Liang, Zili Lin, Haowen Hou, Yong-Lu Li, and Cewu Lu. ImDy: Human inverse dynamics from imitated observations. In International Conference on Learning Representations, 2025.arxiv.org/abs/2410.17610
[29]Liu, X., Ahm, J., Lobaton, E., Si, J., & Huang, H. (2026). Imitation learning from human motion alone does not guarantee biomechanically plausible gait kinetics. In arXiv preprint arXiv:2603.12408. 2603.12408
[30]Ian D Loram and Martin Lakie. Direct measurement of human ankle stiffness during quiet standing: the intrinsic mechanical stiffness is insufficient for stability. The journal of physiology, 545(3):1041–1053, 2002.pubmed.ncbi.nlm.nih.gov/12482906
[31]Zhengyi Luo, Jinkun Cao, Alexander Winkler, Kris Kitani, and Weipeng Xu. Perpetual humanoid control for real-time simulated avatars. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 10895–10904, 2023.arxiv.org/abs/2305.06456
[34]Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. AMASS: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pp. 5442–5451, 2019.arxiv.org/abs/1904.03278
[35]Roberto Martín-Martín, Michelle A Lee, Rachel Gardner, Silvio Savarese, Jeannette Bohg, and Animesh Garg. Variable impedance control in end-effector space: An action space for reinforcement learning in contact-rich tasks. In 2019 IEEE/RSJ international conference on intelligent robots and systems (IROS), pp. 1010–1017. IEEE, 2019.arxiv.org/abs/1906.08880
[36]Thomas A McMahon and George C Cheng. The mechanics of running: how does stiffness couple with speed? Journal of biomechanics, 23:65–78, 1990.pubmed.ncbi.nlm.nih.gov/2081746
[37]Matthew Millard, Thomas Uchida, Ajay Seth, and Scott L Delp. Flexing computational muscle: modeling and simulation of musculotendon dynamics. Journal of biomechanical engineering, 135(2):021005, 2013.DOI
[38]Jason K Moore, Sandra K Hnat, and Antonie J van den Bogert. An elaborate data set on human gait and the effect of mechanical perturbations. PeerJ, 3:e918, 2015.DOI
[39]Lucas Mourot, Ludovic Hoyet, François Le Clerc, and Pierre Hellier. UnderPressure: Deep learning for foot contact detection, ground reaction force estimation and footskate cleanup. In Computer Graphics Forum, volume 41, pp. 195–206. Wiley Online Library, 2022.arxiv.org/abs/2208.04598
[40]Peng, X. B., & Van De Panne, M. (2017). Learning locomotion skills using DeepRL: Does the choice of action space matter?. Proceedings of the ACM SIGGRAPH/Eurographics Symposium on Computer Animation, 1–13.
[41]Peng, X. B., Abbeel, P., Levine, S., & Van de Panne, M. (2018). DeepMimic: Example-guided deep reinforcement learning of physics-based character skills. ACM Transactions On Graphics (TOG), 37(4), 1–14.
[42]Peng, X. B., Ma, Z., Abbeel, P., Levine, S., & Kanazawa, A. (2021). AMP: Adversarial motion priors for stylized physics-based character control. ACM Transactions on Graphics (ToG), 40(4), 1–20.
[43]Lerrel Pinto, Marcin Andrychowicz, Peter Welinder, Wojciech Zaremba, and Pieter Abbeel. Asymmetric actor critic for image-based robot learning. arXiv preprint arXiv:1710.06542, 2017.arXiv
[44]Apoorva Rajagopal, Christopher L Dembia, Matthew S DeMers, Denny D Delp, Jennifer L Hicks, and Scott L Delp. Full-body musculoskeletal model for muscle-driven simulation of human gait. IEEE transactions on biomedical engineering, 63(10):2068–2079, 2016.DOI
[45]Elliott J Rouse, Levi J Hargrove, Eric J Perreault, and Todd A Kuiken. Estimation of human ankle impedance during the stance phase of walking. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 22(4):870–878, 2014.DOI
[46]Keaton Scherpereel, Dean Molinaro, Omer Inan, Max Shepherd, and Aaron Young. A human lower-limb biomechanics and wearable sensors dataset during cyclic and non-cyclic activities. Scientific Data, 10(1):924, 2023.DOI
[47]John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017.arXiv
[48]Shimada, S., Golyanik, V., Xu, W., Pérez, P., & Theobalt, C. (2021). Neural monocular 3D human motion capture with physical awareness. ACM Transactions on Graphics (ToG), 40(4), 1–15.
[49]Merkourios Simos, Alberto Silvio Chiappa, and Alexander Mathis. KINESIS: Motion imitation for human musculoskeletal locomotion. In 2026 IEEE International Conference on Robotics and Automation (ICRA), pp. 20907–20914. IEEE, 2026.arxiv.org/abs/2503.14637
[50]Frank Sup, Amit Bohara, and Michael Goldfarb. Design and control of a powered transfemoral prosthesis. The International journal of robotics research, 27(2):263–273, 2008.pubmed.ncbi.nlm.nih.gov/19898683
[51]Jie Tan, Karen Liu, and Greg Turk. Stable proportional-derivative controllers. IEEE Computer Graphics and Applications, 31(4):34–44, 2011.DOI
[53]Tian Tan, Zachary A Strout, Roy TH Cheung, and Peter B Shull. Strike index estimation using a convolutional neural network with a single, shoe-mounted inertial sensor. Journal of Biomechanics, 139:111145, 2022.DOI
[54]Tian Tan, Dianxin Wang, Peter B Shull, and Eni Halilaj. IMU and smartphone camera fusion for knee adduction and knee flexion moment estimation during walking. IEEE Transactions on Industrial Informatics, 19(2):1445–1455, 2023.DOI
[55]Tian Tan, Tom Van Wouwe, Keenon F Werling, C Karen Liu, Scott L Delp, Jennifer L Hicks, and Akshay S Chaudhari. GaitDynamics: a generative foundation model for analyzing human walking and running. Nature Biomedical Engineering, 10:1659–1671, 2026.DOI
[56]Tessler, C., Guo, Y., Nabati, O., Chechik, G., & Peng, X. B. (2024). MaskedMimic: Unified physics-based character control through masked motion inpainting. ACM Transactions On Graphics (TOG), 43(6), 1–21.
[57]Emanuel Todorov, Tom Erez, and Yuval Tassa. MuJoCo: A physics engine for model-based control. In 2012 IEEE/RSJ international conference on intelligent robots and systems, pp. 5026–5033. IEEE, 2012.DOI
[58]Huawei Wang, Vittorio Caggiano, Guillaume Durandau, Massimo Sartori, and Vikash Kumar. MyoSim: Fast and physiologically realistic MuJoCo models for musculoskeletal and exoskeletal studies. In 2022 International Conference on Robotics and Automation (ICRA), pp. 8104–8111. IEEE, 2022.DOI
[59]Huawei Wang, Akash Basu, Guillaume Durandau, and Massimo Sartori. A wearable real-time kinetic measurement sensor setup for human locomotion. Wearable technologies, 4:e11, 2023.DOI
[60]Keenon Werling, Nicholas A Bianco, Michael Raitor, Jon Stingel, Jennifer L Hicks, Steven H Collins, Scott L Delp, and C Karen Liu. AddBiomechanics: Automating model scaling, inverse kinematics, and inverse dynamics from human motion data through sequential optimization. Plos one, 18(11):e0295152, 2023.DOI
[61]Werling, K., Kaneda, J., Tan, T., Agarwal, R., Skov, S., Van Wouwe, T., Uhlrich, S., Bianco, N., Ong, C., Falisse, A., & al., . et . (2024). AddBiomechanics dataset: Capturing the physics of human motion at scale. European Conference on Computer Vision, 490–508.
[62]Ye Yuan and Kris Kitani. Residual force control for agile human behavior imitation and extended motion synthesis. Advances in Neural Information Processing Systems, 33:21763–21774, 2020.arxiv.org/abs/2006.07364
[63]Ye Yuan, Shih-En Wei, Tomas Simon, Kris Kitani, and Jason Saragih. SimPoE: Simulated character control for 3D human pose estimation. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 7155–7165. IEEE, 2021.arxiv.org/abs/2104.00683
[64]Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. PhysDiff: Physics-guided human motion diffusion model. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 15964–15975. IEEE, 2023.arxiv.org/abs/2212.02500
[65]Zhang, J., Liu, J., Lee, Y.-Y., Moon, S., Zordan, V., Tevet, G., Liu, C. K., Gould, S., Jacob, O., Jiang, H., & RoMo, A. (2026). RoMo: A large-scale, richly organized dataset and semantic taxonomy for human motion generation. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 16408–16419.