Manoj Srinivasan
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Research Interests
Manoj Srinivasan has a total of 38 co-authors. Here are the top co-authors by number of shared publications:
- Andy Ruina
Cornell University
5 shared publications - Nidhi Seethapathi
Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology, Cambridge, MA, USA
5 shared publications - Aya Alwan
Mechanical and Aerospace Engineering, The Ohio State University, Columbus, OH 43210, USA
4 shared publications - Matthew L. Handford
Department of Mechanical and Aerospace Engineering, The Ohio State University, Columbus, OH, USA
3 shared publications - Varun Joshi
School of Kinesiology, University of Michigan, Ann Arbor, Michigan, United States of America
3 shared publications Yang Wang
Mechanical and Aerospace Engineering, The Ohio State University, Columbus, OH 43210, USA
3 shared publications
Latest Publications
- Open Access
Highly under-actuated dynamic manipulation: Dice stacking is mostly open-loop
Humans’ ability to grasp and dynamically manipulate objects with their hands is unmatched by current robots. To better understand human dynamic manipulation, we studied dice stacking, a task in which humans form a vertical stack of dice from a set of initially unstacked playing dice using an overturned cup and the surface of a table. This task is high dimensional and under-actuated, so it may superficially seem an incredible feat of state estimation and feedback control, but we show that this task is amenable to open-loop strategies. We simulated a cup with dice oscillated by fixed arm movement patterns using two different computer simulation frameworks with different contact models. These simulations showed that, for a range of arm and wrist movements, the dice naturally stack without any dice state feedback. We verified the predictions of these simulations with a physical robot. Thus, we have added dice stacking to the small list of dynamic manipulation tasks that can be robustly performed open-loop. We speculate that, for highly under-actuated tasks, humans may be biased to learn open-loop strategies over state feedback strategies. Future work could investigate the presence of such a bias in humans and its potential value for reinforcement learning algorithms.
Open AccessHighly under-actuated dynamic manipulation: Dice stacking is mostly open-loop
Humans’ ability to grasp and dynamically manipulate objects with their hands is unmatched by current robots. To better understand human dynamic manipulation, we studied dice stacking, a task in which humans form a vertical stack of dice from a set of initially unstacked playing dice using an overturned cup and the surface of a table. This task is high dimensional and under-actuated, so it may superficially seem an incredible feat of state estimation and feedback control, but we show that this task is amenable to open-loop strategies. We simulated a cup with dice oscillated by fixed arm movement patterns using two different computer simulation frameworks with different contact models. These simulations showed that, for a range of arm and wrist movements, the dice naturally stack without any dice state feedback. We verified the predictions of these simulations with a physical robot. Thus, we have added dice stacking to the small list of dynamic manipulation tasks that can be robustly performed open-loop. We speculate that, for highly under-actuated tasks, humans may be biased to learn open-loop strategies over state feedback strategies. Future work could investigate the presence of such a bias in humans and its potential value for reinforcement learning algorithms.
- Open Access
Paradoxical noise preference in RNNs
In recurrent neural networks (RNNs) used to model biological neural networks, noise is typically introduced during training to emulate biological variability and regularize learning. The expectation is that removing the noise at test time should preserve or improve performance. Contrary to this intuition, we find that continuous-time recurrent neural networks (CTRNNs) often perform best at a nonzero noise level, specifically, the same level used during training. This noise preference typically arises when noise is injected inside the neural activation function; networks trained with noise injected outside the activation function perform best with zero noise. Through analyses of simple function approximation, maze navigation, and single neuron regulator tasks, we show that the phenomenon stems from noise-induced shifts of fixed points (stationary distributions) in the underlying stochastic dynamics of the RNNs. These fixed point shifts are noise-level dependent and bias the network outputs when the noise is removed, degrading performance. Analytical and numerical results show that the bias arises when neural states operate near activation function nonlinearities, where noise is asymmetrically attenuated, and that performance optimization incentivizes operation near these nonlinearities. Thus, networks can overfit to the stochastic training environment itself rather than just to the input-output data. The phenomenon is distinct from stochastic resonance, wherein nonzero noise enhances signal processing. Our findings reveal that training noise can become an integral part of the computation learned by recurrent networks, with implications for understanding neural population dynamics and for the design of robust artificial RNNs.
Open AccessParadoxical noise preference in RNNs
In recurrent neural networks (RNNs) used to model biological neural networks, noise is typically introduced during training to emulate biological variability and regularize learning. The expectation is that removing the noise at test time should preserve or improve performance. Contrary to this intuition, we find that continuous-time recurrent neural networks (CTRNNs) often perform best at a nonzero noise level, specifically, the same level used during training. This noise preference typically arises when noise is injected inside the neural activation function; networks trained with noise injected outside the activation function perform best with zero noise. Through analyses of simple function approximation, maze navigation, and single neuron regulator tasks, we show that the phenomenon stems from noise-induced shifts of fixed points (stationary distributions) in the underlying stochastic dynamics of the RNNs. These fixed point shifts are noise-level dependent and bias the network outputs when the noise is removed, degrading performance. Analytical and numerical results show that the bias arises when neural states operate near activation function nonlinearities, where noise is asymmetrically attenuated, and that performance optimization incentivizes operation near these nonlinearities. Thus, networks can overfit to the stochastic training environment itself rather than just to the input-output data. The phenomenon is distinct from stochastic resonance, wherein nonzero noise enhances signal processing. Our findings reveal that training noise can become an integral part of the computation learned by recurrent networks, with implications for understanding neural population dynamics and for the design of robust artificial RNNs.
- Open Access
- 5 November 2025
- Published by The Royal Society in Royal Society Open Science
- Vol. 12 (11)
Overground gait transitions are not sharp but involve gradually changing walk–run mixtures even over long distances
Humans typically walk at low speeds and run at higher speeds. Previous studies of transitions between walking and running were mostly on treadmills, but real-world locomotion allows more flexibility. Here, we study overground locomotion over long distances (800 or 2400 m) under time constraints, simulating everyday scenarios like traveling to an appointment. Unlike on treadmills, participants can vary both speed and gait during this task. Gait transition in this overground task occurs over a broad ‘gait transition regime’ spanning average speeds from 1.9 to 3.0 m s −1 . In this regime, people use mixtures of walking and running on each travel bout: mostly walking at low average speeds (around 1.9 m s −1 ) and mostly running at high average speeds (3.0 m s −1 ). The walk–run fraction changes gradually between these speed limits and is 50% at about 2.5 m s −1 . Within each walk–run mixture, walking is slower than running, with an unused gap between the two gait speeds. These gait mixtures and their speed dependence are predicted by energy optimality. These findings extend earlier results for shorter distances, showing that similar energetic principles govern longer, more physically and cognitively demanding tasks. Our results highlight the role of whole-task energy minimization including transients in shaping human locomotion and gait choice.
- 5 November 2025
- Published by The Royal Society in Royal Society Open Science
- Vol. 12 (11)
Open AccessOverground gait transitions are not sharp but involve gradually changing walk–run mixtures even over long distances
Humans typically walk at low speeds and run at higher speeds. Previous studies of transitions between walking and running were mostly on treadmills, but real-world locomotion allows more flexibility. Here, we study overground locomotion over long distances (800 or 2400 m) under time constraints, simulating everyday scenarios like traveling to an appointment. Unlike on treadmills, participants can vary both speed and gait during this task. Gait transition in this overground task occurs over a broad ‘gait transition regime’ spanning average speeds from 1.9 to 3.0 m s −1 . In this regime, people use mixtures of walking and running on each travel bout: mostly walking at low average speeds (around 1.9 m s −1 ) and mostly running at high average speeds (3.0 m s −1 ). The walk–run fraction changes gradually between these speed limits and is 50% at about 2.5 m s −1 . Within each walk–run mixture, walking is slower than running, with an unused gap between the two gait speeds. These gait mixtures and their speed dependence are predicted by energy optimality. These findings extend earlier results for shorter distances, showing that similar energetic principles govern longer, more physically and cognitively demanding tasks. Our results highlight the role of whole-task energy minimization including transients in shaping human locomotion and gait choice.
- Open Access
- 5 November 2025
- Published by The Royal Society in Journal of The Royal Society Interface
- Vol. 22 (232) , 20250602
Natural variability can increase human walking metabolic costs and its implications to simulation-based metabolic estimation
Human walking contains variability due to small intrinsic perturbations arising from sensory or motor noise, or to promote motor learning. We hypothesize that such stride-to-stride variability may increase the metabolic cost of walking over and above a perfectly periodic motion, and that neglecting such variability in simulations may mis-estimate the metabolic cost. Here, we quantify the metabolic estimation errors accrued by neglecting the stride-to-stride variability using human data and a musculoskeletal model by comparing the cost of multiple strides of walking and the cost of a perfectly periodic stride with averaged kinematics and kinetics. We find that using an averaged stride underestimates the cost by approximately 2.5%, whereas using a random stride may mis-estimate the cost positively or negatively by up to 15%, ignoring the contribution of measurement errors to the observed stride-to-stride variability. As a further illustration of the cost increase in a simpler dynamical context, we use a feedback-controlled inverted pendulum walking model to show that increasing the sensory or motor noise increases the overall metabolic cost, as well as the variability of stride-to-stride metabolic costs, as seen with the musculoskeletal simulations. Our work establishes the importance of accounting for stride-to-stride variability when estimating metabolic costs from motion.
- 5 November 2025
- Published by The Royal Society in Journal of The Royal Society Interface
- Vol. 22 (232) , 20250602
Open AccessNatural variability can increase human walking metabolic costs and its implications to simulation-based metabolic estimation
Human walking contains variability due to small intrinsic perturbations arising from sensory or motor noise, or to promote motor learning. We hypothesize that such stride-to-stride variability may increase the metabolic cost of walking over and above a perfectly periodic motion, and that neglecting such variability in simulations may mis-estimate the metabolic cost. Here, we quantify the metabolic estimation errors accrued by neglecting the stride-to-stride variability using human data and a musculoskeletal model by comparing the cost of multiple strides of walking and the cost of a perfectly periodic stride with averaged kinematics and kinetics. We find that using an averaged stride underestimates the cost by approximately 2.5%, whereas using a random stride may mis-estimate the cost positively or negatively by up to 15%, ignoring the contribution of measurement errors to the observed stride-to-stride variability. As a further illustration of the cost increase in a simpler dynamical context, we use a feedback-controlled inverted pendulum walking model to show that increasing the sensory or motor noise increases the overall metabolic cost, as well as the variability of stride-to-stride metabolic costs, as seen with the musculoskeletal simulations. Our work establishes the importance of accounting for stride-to-stride variability when estimating metabolic costs from motion.
- Open Access
- 16 April 2025
- Published by Springer Nature in Annals of Biomedical Engineering
- Vol. 53 (7) , 1604-1614
James–Stein Estimator Improves Accuracy and Sample Efficiency in Human Kinematic and Metabolic Data
Human biomechanical data are often accompanied with measurement noise and behavioral variability. Errors due to such noise and variability are usually exaggerated by fewer trials or shorter trial durations and could be reduced using more trials or longer trial durations. Speeding up such data collection by lowering number of trials or trial duration, while improving the accuracy of statistical estimates, would be of particular interest in wearable robotics applications and when the human population studied is vulnerable (e.g., the elderly). Here, we propose the use of the James–Stein estimator (JSE) to improve statistical estimates with a given amount of data or reduce the amount of data needed for a given accuracy. The JSE is a shrinkage estimator that produces a uniform reduction in the summed squared errors when compared with the more familiar maximum likelihood estimator (MLE), simple averages, or other least squares regressions. When data from multiple human participants are available, an individual participant’s JSE can improve upon MLE by incorporating information from all participants, improving overall estimation accuracy on average. Here, we apply the JSE to multiple time series of kinematic and metabolic data from the following parameter estimation problems: foot placement control during level walking, energy expenditure during circle walking, and energy expenditure during resting. We show that the resulting estimates improve accuracy—that is, the James–Stein estimates have lower summed squared error from the ‘true’ value compared with more conventional estimates.
- 16 April 2025
- Published by Springer Nature in Annals of Biomedical Engineering
- Vol. 53 (7) , 1604-1614
Open AccessJames–Stein Estimator Improves Accuracy and Sample Efficiency in Human Kinematic and Metabolic Data
Human biomechanical data are often accompanied with measurement noise and behavioral variability. Errors due to such noise and variability are usually exaggerated by fewer trials or shorter trial durations and could be reduced using more trials or longer trial durations. Speeding up such data collection by lowering number of trials or trial duration, while improving the accuracy of statistical estimates, would be of particular interest in wearable robotics applications and when the human population studied is vulnerable (e.g., the elderly). Here, we propose the use of the James–Stein estimator (JSE) to improve statistical estimates with a given amount of data or reduce the amount of data needed for a given accuracy. The JSE is a shrinkage estimator that produces a uniform reduction in the summed squared errors when compared with the more familiar maximum likelihood estimator (MLE), simple averages, or other least squares regressions. When data from multiple human participants are available, an individual participant’s JSE can improve upon MLE by incorporating information from all participants, improving overall estimation accuracy on average. Here, we apply the JSE to multiple time series of kinematic and metabolic data from the following parameter estimation problems: foot placement control during level walking, energy expenditure during circle walking, and energy expenditure during resting. We show that the resulting estimates improve accuracy—that is, the James–Stein estimates have lower summed squared error from the ‘true’ value compared with more conventional estimates.
- Open Access
Natural variability increases human walking metabolic costs and its implications to simulation-based metabolic estimation
Human walking contains variability due to small intrinsic perturbations arising from sensory or motor noise, or to promote motor learning. We hypothesize that such stride-to-stride variability may increase the metabolic cost of walking over and above a perfectly periodic motion, and that neglecting such variability in simulations may mis-estimate the metabolic cost. Here, we quantify the metabolic estimation errors accrued by neglecting the stride-to-stride variability using human data and a musculoskeletal model by comparing the cost of multiple strides of walking and the cost of a perfectly periodic stride with averaged kinematics and kinetics. We find that using an averaged stride underestimates the cost by about 2.5%, whereas using a random stride may mis-estimate the cost positively or negatively by up to 15%. As a further illustration of the cost increase in a simpler dynamical context, we use a feedback-controlled inverted pendulum walking model to show that increasing the sensory or motor noise increases the overall metabolic cost, as well as the variability of stride-to-stride metabolic costs, as seen with the musculoskeletal simulations. Our work establishes the importance of accounting for stride-to-stride variability when estimating metabolic costs from motion.
Open AccessNatural variability increases human walking metabolic costs and its implications to simulation-based metabolic estimation
Human walking contains variability due to small intrinsic perturbations arising from sensory or motor noise, or to promote motor learning. We hypothesize that such stride-to-stride variability may increase the metabolic cost of walking over and above a perfectly periodic motion, and that neglecting such variability in simulations may mis-estimate the metabolic cost. Here, we quantify the metabolic estimation errors accrued by neglecting the stride-to-stride variability using human data and a musculoskeletal model by comparing the cost of multiple strides of walking and the cost of a perfectly periodic stride with averaged kinematics and kinetics. We find that using an averaged stride underestimates the cost by about 2.5%, whereas using a random stride may mis-estimate the cost positively or negatively by up to 15%. As a further illustration of the cost increase in a simpler dynamical context, we use a feedback-controlled inverted pendulum walking model to show that increasing the sensory or motor noise increases the overall metabolic cost, as well as the variability of stride-to-stride metabolic costs, as seen with the musculoskeletal simulations. Our work establishes the importance of accounting for stride-to-stride variability when estimating metabolic costs from motion.
- Open Access
- 3 November 2024
- Published by Springer Nature in Nature Communications
- Vol. 15 (1) , 1-23
Exploration-based learning of a stabilizing controller predicts locomotor adaptation
Humans adapt their locomotion seamlessly in response to changes in the body or the environment. It is unclear how such adaptation improves performance measures like energy consumption or symmetry while avoiding falling. Here, we model locomotor adaptation as interactions between a stabilizing controller that reacts quickly to perturbations and a reinforcement learner that gradually improves the controller’s performance through local exploration and memory. This model predicts time-varying adaptation in many settings: walking on a split-belt treadmill (i.e. with both feet at different speeds), with asymmetric leg weights, or using exoskeletons — capturing learning and generalization phenomena in ten prior experiments and two model-guided experiments conducted here. The performance measure of energy minimization with a minor cost for asymmetry captures a broad range of phenomena and can act alongside other mechanisms such as reducing sensory prediction error. Such a model-based understanding of adaptation can guide rehabilitation and wearable robot control.
- 3 November 2024
- Published by Springer Nature in Nature Communications
- Vol. 15 (1) , 1-23
Open AccessExploration-based learning of a stabilizing controller predicts locomotor adaptation
Humans adapt their locomotion seamlessly in response to changes in the body or the environment. It is unclear how such adaptation improves performance measures like energy consumption or symmetry while avoiding falling. Here, we model locomotor adaptation as interactions between a stabilizing controller that reacts quickly to perturbations and a reinforcement learner that gradually improves the controller’s performance through local exploration and memory. This model predicts time-varying adaptation in many settings: walking on a split-belt treadmill (i.e. with both feet at different speeds), with asymmetric leg weights, or using exoskeletons — capturing learning and generalization phenomena in ten prior experiments and two model-guided experiments conducted here. The performance measure of energy minimization with a minor cost for asymmetry captures a broad range of phenomena and can act alongside other mechanisms such as reducing sensory prediction error. Such a model-based understanding of adaptation can guide rehabilitation and wearable robot control.
- Open Access
James-Stein estimator improves accuracy and sample efficiency in human kinematic and metabolic data
Human biomechanical data are often accompanied with measurement noise and behavioral variability. Errors due to such noise and variability are usually exaggerated by fewer trials or shorter trial durations, and could be reduced using more trials or longer trial durations. Speeding up such data collection by lowering number of trials or trial duration, while improving the accuracy of statistical estimates, would be of particular interest in wearable robotics applications and when the human population studied is vulnerable (e.g., the elderly). Here, we propose the use of the James-Stein estimator (JSE) to improve statistical estimates with a given amount of data, or reduce the amount of data needed for a given accuracy. The JSE is a shrinkage estimator that produces a uniform reduction in the summed squared errors when compared with the more familiar maximum likelihood estimator (MLE), simple averages, or other least squares regressions. When data from multiple human participants are available, an individual participant’s JSE can improve upon MLE by incorporating information from all participants, improving overall estimation accuracy on average. Here, we apply the JSE to multiple time-series of kinematic and metabolic data from the following parameter estimation problems: foot placement control during level walking, energy expenditure during circle walking, and energy expenditure during resting. We show that the resulting estimates improve accuracy — that is, the James-Stein estimates have lower summed squared error from the ‘true’ value compared with more conventional estimates.
Open AccessJames-Stein estimator improves accuracy and sample efficiency in human kinematic and metabolic data
Human biomechanical data are often accompanied with measurement noise and behavioral variability. Errors due to such noise and variability are usually exaggerated by fewer trials or shorter trial durations, and could be reduced using more trials or longer trial durations. Speeding up such data collection by lowering number of trials or trial duration, while improving the accuracy of statistical estimates, would be of particular interest in wearable robotics applications and when the human population studied is vulnerable (e.g., the elderly). Here, we propose the use of the James-Stein estimator (JSE) to improve statistical estimates with a given amount of data, or reduce the amount of data needed for a given accuracy. The JSE is a shrinkage estimator that produces a uniform reduction in the summed squared errors when compared with the more familiar maximum likelihood estimator (MLE), simple averages, or other least squares regressions. When data from multiple human participants are available, an individual participant’s JSE can improve upon MLE by incorporating information from all participants, improving overall estimation accuracy on average. Here, we apply the JSE to multiple time-series of kinematic and metabolic data from the following parameter estimation problems: foot placement control during level walking, energy expenditure during circle walking, and energy expenditure during resting. We show that the resulting estimates improve accuracy — that is, the James-Stein estimates have lower summed squared error from the ‘true’ value compared with more conventional estimates.
- Open Access
- 2 January 2024
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 19 (1) , e0295993
Walking speeds are lower for short distance and turning locomotion: Experiments and modeling in low-cost prosthesis users
Preferred walking speed is a widely-used performance measure for people with mobility issues, but is usually measured in straight line walking for fixed distances or durations, and without explicitly accounting for turning. However, daily walking involves walking for bouts of different distances and walking with turning, with prior studies showing that short bouts with at most 10 steps could be 40% of all bouts and turning steps could be 8-50% of all steps. Here, we studied walking in a straight line for short distances (4 m to 23 m) and walking in circles (1 m to 3 m turning radii) in people with transtibial amputation or transfemoral amputation using a passive ankle-foot prosthesis (Jaipur Foot). We found that the study participants’ preferred walking speeds are lower for shorter straight-line walking distances and lower for circles of smaller radii, which is analogous to earlier results in subjects without amputation. Using inverse optimization, we estimated the cost of changing speeds and turning such that the observed preferred walking speeds in our experiments minimizes the total cost of walking. The inferred costs of changing speeds and turning were larger for subjects with amputation compared to subjects without amputation in a previous study, specifically, being 4x to 8x larger for the turning cost and being highest for subjects with transfemoral amputation. Such high costs inferred by inverse optimization could potentially include non-energetic costs such as due to joint or interfacial stress or stability concerns, as inverse optimization cannot distinguish such terms from true metabolic cost. These experimental findings and models capturing the experimental trends could inform prosthesis design and rehabilitation therapy to better assist changing speeds and turning tasks. Further, measuring the preferred speed for a range of distances and radii could be a more comprehensive subject-specific measure of walking performance than commonly used straight line walking metrics.
- 2 January 2024
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 19 (1) , e0295993
Open AccessWalking speeds are lower for short distance and turning locomotion: Experiments and modeling in low-cost prosthesis users
Preferred walking speed is a widely-used performance measure for people with mobility issues, but is usually measured in straight line walking for fixed distances or durations, and without explicitly accounting for turning. However, daily walking involves walking for bouts of different distances and walking with turning, with prior studies showing that short bouts with at most 10 steps could be 40% of all bouts and turning steps could be 8-50% of all steps. Here, we studied walking in a straight line for short distances (4 m to 23 m) and walking in circles (1 m to 3 m turning radii) in people with transtibial amputation or transfemoral amputation using a passive ankle-foot prosthesis (Jaipur Foot). We found that the study participants’ preferred walking speeds are lower for shorter straight-line walking distances and lower for circles of smaller radii, which is analogous to earlier results in subjects without amputation. Using inverse optimization, we estimated the cost of changing speeds and turning such that the observed preferred walking speeds in our experiments minimizes the total cost of walking. The inferred costs of changing speeds and turning were larger for subjects with amputation compared to subjects without amputation in a previous study, specifically, being 4x to 8x larger for the turning cost and being highest for subjects with transfemoral amputation. Such high costs inferred by inverse optimization could potentially include non-energetic costs such as due to joint or interfacial stress or stability concerns, as inverse optimization cannot distinguish such terms from true metabolic cost. These experimental findings and models capturing the experimental trends could inform prosthesis design and rehabilitation therapy to better assist changing speeds and turning tasks. Further, measuring the preferred speed for a range of distances and radii could be a more comprehensive subject-specific measure of walking performance than commonly used straight line walking metrics.
- Open Access
Metabolic cost for isometric force scales nonlinearly and predicts how humans distribute forces across limbs
Muscles consume metabolic energy for force production and movement. A mathematical model of metabolic energy cost will be useful in predicting instantaneous costs during human exercise and in computing effort-minimizing movements via simulations. Previous in vivo data-derived models usually assumed either zero or linearly increasing cost with force, but a nonlinear relation could have significant metabolic or behavioural implications. Here, we show that metabolic cost scales nonlinearly with joint torque with an exponent of about 1.64, using calorimetric measurements of isometric squats. We then demonstrate that this metabolic nonlinearity is reflected in human behaviour: minimizing this nonlinear cost predicts how humans share forces between limbs in additional experiments involving arms and legs. This shows the utility of the nonlinear energy cost in predictive models and its generalizability across limbs. Finally, we show mathematical evidence that the same nonlinear metabolic objective may underlie force sharing at the muscle level.
Open AccessMetabolic cost for isometric force scales nonlinearly and predicts how humans distribute forces across limbs
Muscles consume metabolic energy for force production and movement. A mathematical model of metabolic energy cost will be useful in predicting instantaneous costs during human exercise and in computing effort-minimizing movements via simulations. Previous in vivo data-derived models usually assumed either zero or linearly increasing cost with force, but a nonlinear relation could have significant metabolic or behavioural implications. Here, we show that metabolic cost scales nonlinearly with joint torque with an exponent of about 1.64, using calorimetric measurements of isometric squats. We then demonstrate that this metabolic nonlinearity is reflected in human behaviour: minimizing this nonlinear cost predicts how humans share forces between limbs in additional experiments involving arms and legs. This shows the utility of the nonlinear energy cost in predictive models and its generalizability across limbs. Finally, we show mathematical evidence that the same nonlinear metabolic objective may underlie force sharing at the muscle level.