Constraint-Consistent Torque Controller

• Constraint-consistent torque controllers are control laws that embed task, safety, kinematic, and actuator constraints directly into the torque synthesis process.
• Key synthesis methods include probabilistic fusion, convex optimization, projection-based inverse dynamics, and predictive feedforward optimization.
• Practical implementations achieve improved tracking, reduced oscillations, and real-time performance in diverse platforms such as manipulators, bipeds, and surgical robots.

Searching arXiv for recent and foundational papers on constraint-consistent torque control and closely related formulations. Across these works, a constraint-consistent torque controller is a torque-level control law in which the commanded torque is synthesized together with task, safety, kinematic, force, contact, or actuator constraints rather than after them. The common objective is to preserve consistency between the commanded torque and the constrained dynamics of the underlying system, whether the system is a torque-controlled manipulator, a biped, a surgical robot, a synchronous machine, or a mobile platform. Representative formulations include probabilistic fusion of elementary torque laws, quadratic programs with Lyapunov or barrier constraints, projection-based inverse dynamics for holonomic constraints, viability-based safe torque projection, predictive optimization under input and state limits, and analytic feedforward solutions under current and voltage constraints (Silvério et al., 2017, Galloway et al., 2013, Zhang et al., 2024, Li et al., 17 Sep 2025, Stumper et al., 2012, Eldeeb et al., 2016).

1. Definition and dynamical setting

A recurrent starting point is the rigid-body dynamics of a fully actuated manipulator,

$M(q)\,\ddot q + C(q,\dot q)\,\dot q + G(q) = \tau,$

or closely related notations such as

$$H(q)\,\ddot q + C(q,\dot q)\,\dot q + G(q)=\tau_c+\tau_{ext},$$

and, for bipedal systems,

$$D(q)\,\ddot q + C(q,\dot q)\,\dot q + G(q)=B(q)\,u.$$

In electrical-drive applications the same design idea appears in linearized state-space form, for example

$\dot x(t)=A\,x(t)+B\,u(t)+d,$

with explicit input and state constraints (Pasandi et al., 2023, Zhang et al., 2024, Galloway et al., 2013, Stumper et al., 2012).

Within this setting, “constraint-consistent” denotes a controller whose torque command is computed so that the constraint representation remains compatible with the system dynamics. Depending on the domain, the relevant constraints are expressed as task Jacobian relations, force-tracking laws, torque saturations, current and voltage bounds, control Lyapunov inequalities, control barrier inequalities, viability conditions, friction-cone conditions, unilateral-contact conditions, or rheonomic holonomic constraints such as remote center of motion (RCM) constraints in surgical robotics (Silvério et al., 2017, Dafarra et al., 2017, Li et al., 17 Sep 2025).

This suggests that the topic is not a single algorithmic family but a design principle. The principle is to move the constraint representation into the torque-synthesis stage itself, rather than treating torque generation and constraint handling as separate layers.

2. Main synthesis mechanisms

One major mechanism is probabilistic torque fusion. In “Probabilistic Learning of Torque Controllers from Kinematic and Force Constraints,” each elementary torque controller $i$ defines a Gaussian torque distribution

$p_i(\tau)=\mathcal{N}(\tau;\mu_i,\Sigma_i),$

and the fused torque is obtained from the product of Gaussians,

$$\tau^*=\mu^*=\Bigl(\sum_{i=1}^N \Sigma_i^{-1}\Bigr)^{-1}\sum_{i=1}^N \Sigma_i^{-1}\mu_i.$$

The covariance encodes controller relevance or confidence, so small eigenvalues of $\Sigma_i$ force accurate satisfaction along the associated torque directions (Silvério et al., 2017).

A second mechanism is torque synthesis by convex optimization. In CLF-based walking control, the pointwise min-norm auxiliary input is written as

$$\min_\mu \ \mu^T\mu \quad\text{s.t.}\quad \psi_0+\psi_1\,\mu\le 0,$$

and torque saturation is added as linear constraints on $\mu$, either with hard bounds or with slack variables $$H(q)\,\ddot q + C(q,\dot q)\,\dot q + G(q)=\tau_c+\tau_{ext},$$0 that trade off CLF decay against saturation compliance (Galloway et al., 2013). In constrained passive interaction control and safety-critical adaptive impedance control, the QP objective minimizes deviation from a nominal passive or impedance-based command while hard or soft constraints encode joint limits, collisions, singularities, and torque limits (Zhang et al., 2024, Lawan et al., 27 May 2026).

A third mechanism is projection-based inverse dynamics under explicit constraints. For surgical RCM control, the constraint projector

$$H(q)\,\ddot q + C(q,\dot q)\,\dot q + G(q)=\tau_c+\tau_{ext},$$1

splits the dynamics into free and constrained subspaces. The torque is then decomposed as

$$H(q)\,\ddot q + C(q,\dot q)\,\dot q + G(q)=\tau_c+\tau_{ext},$$2

with $$H(q)\,\ddot q + C(q,\dot q)\,\dot q + G(q)=\tau_c+\tau_{ext},$$3 for free-motion tasks and $$H(q)\,\ddot q + C(q,\dot q)\,\dot q + G(q)=\tau_c+\tau_{ext},$$4 for constraint enforcement. In that construction, the RCM is treated as a rheonomic holonomic constraint embedded directly into inverse dynamics (Li et al., 17 Sep 2025).

A fourth mechanism is explicit predictive or feedforward constrained optimization. For linear control systems with input and state constraints, flatness-based trajectory generation yields a finite-dimensional QP or LP whose solution generates a predictive torque controller for a permanent magnet synchronous motor (Stumper et al., 2012). For anisotropic synchronous machines, optimal feedforward torque control is reformulated as intersections of quadrics, and the solution is obtained analytically from a fourth-order polynomial in the Lagrange multiplier $$H(q)\,\ddot q + C(q,\dot q)\,\dot q + G(q)=\tau_c+\tau_{ext},$$5 (Eldeeb et al., 2016).

Formulation family	Constraint representation	Representative paper
Probabilistic fusion	Gaussian torque distributions and precision-weighted fusion	(Silvério et al., 2017)
CLF/CBF/NCBF QP	Linear inequalities, Lyapunov decay, barrier conditions, slack variables	(Galloway et al., 2013, Zhang et al., 2024, Lawan et al., 27 May 2026)
Projection-based inverse dynamics	Constraint projector and orthogonal torque decomposition	(Li et al., 17 Sep 2025)
Viability-preserving safe control	Acceleration-space inequalities mapped to torque-space	(Zhang et al., 3 Oct 2025)
Predictive or analytic constrained control	Horizon-wise inequalities or quadric/quartic optimality conditions	(Stumper et al., 2012, Eldeeb et al., 2016)

3. Constraint classes and their encoding

Kinematic motion constraints are commonly encoded through task-space Jacobians. In the probabilistic fusion framework, a motion task is written as

$$H(q)\,\ddot q + C(q,\dot q)\,\dot q + G(q)=\tau_c+\tau_{ext},$$6

with $$H(q)\,\ddot q + C(q,\dot q)\,\dot q + G(q)=\tau_c+\tau_{ext},$$7 and covariance chosen through

$$H(q)\,\ddot q + C(q,\dot q)\,\dot q + G(q)=\tau_c+\tau_{ext},$$8

This assigns low priority to torque directions that do not affect task $$H(q)\,\ddot q + C(q,\dot q)\,\dot q + G(q)=\tau_c+\tau_{ext},$$9 and high priority to those that do (Silvério et al., 2017).

Force constraints are encoded similarly. For contact-force maintenance,

$$D(q)\,\ddot q + C(q,\dot q)\,\dot q + G(q)=B(q)\,u.$$0

Smaller entries in $$D(q)\,\ddot q + C(q,\dot q)\,\dot q + G(q)=B(q)\,u.$$1 tighten tracking of the desired force $$D(q)\,\ddot q + C(q,\dot q)\,\dot q + G(q)=B(q)\,u.$$2 (Silvério et al., 2017). In whole-body control of humanoids, desired contact wrenches are optimized under unilateral-contact and Coulomb-friction constraints,

$$D(q)\,\ddot q + C(q,\dot q)\,\dot q + G(q)=B(q)\,u.$$3

and are then mapped into torques through inverse dynamics (Dafarra et al., 2017).

State and actuator limits are handled in several distinct ways. One approach is direct saturation in the torque-synthesis QP:

$$D(q)\,\ddot q + C(q,\dot q)\,\dot q + G(q)=B(q)\,u.$$4

possibly with soft relaxations (Galloway et al., 2013). Another is state transformation. In “Torque Control with Joints Position and Velocity Limits Avoidance,” hard bounds are enforced by introducing exogenous states $$D(q)\,\ddot q + C(q,\dot q)\,\dot q + G(q)=B(q)\,u.$$5 through

$$D(q)\,\ddot q + C(q,\dot q)\,\dot q + G(q)=B(q)\,u.$$6

so bounded exogenous states imply that joint positions and velocities remain strictly inside the feasible open intervals (Pasandi et al., 2023). A third approach is barrier-based: position and velocity limits are encoded as relative-degree-two CBFs or as a composed position-velocity nonsmooth CBF yielding explicit acceleration bounds

$$D(q)\,\ddot q + C(q,\dot q)\,\dot q + G(q)=B(q)\,u.$$7

for each joint (Zhang et al., 2024, Lawan et al., 27 May 2026).

Holonomic and rheonomic constraints form another important class. In surgical robotics, the RCM constraint is written in local tool coordinates as

$$D(q)\,\ddot q + C(q,\dot q)\,\dot q + G(q)=B(q)\,u.$$8

with acceleration-level condition

$$D(q)\,\ddot q + C(q,\dot q)\,\dot q + G(q)=B(q)\,u.$$9

The corresponding torque controller enforces tool-tip tracking while maintaining the trocar constraint in a dynamically consistent manner (Li et al., 17 Sep 2025).

Safety constraints are often encoded as affine inequalities in acceleration and then mapped to torque space. In viability-preserving passive torque control,

$\dot x(t)=A\,x(t)+B\,u(t)+d,$0

is transformed using

$\dot x(t)=A\,x(t)+B\,u(t)+d,$1

into

$\dot x(t)=A\,x(t)+B\,u(t)+d,$2

This construction is used for joint-position limits, joint-velocity limits, self-collision avoidance, external-obstacle avoidance, and hardware acceleration limits (Zhang et al., 3 Oct 2025).

4. Learning, adaptation, and model dependence

Learning from demonstrations provides one route to torque consistency when the relevant task representation is not known in advance. In the Gaussian-fusion framework, demonstrations of $\dot x(t)=A\,x(t)+B\,u(t)+d,$3, $\dot x(t)=A\,x(t)+B\,u(t)+d,$4, and $\dot x(t)=A\,x(t)+B\,u(t)+d,$5 are modeled by a mixture of $\dot x(t)=A\,x(t)+B\,u(t)+d,$6 probabilistic controllers with priors $\dot x(t)=A\,x(t)+B\,u(t)+d,$7. Responsibilities

$\dot x(t)=A\,x(t)+B\,u(t)+d,$8

are computed in the E-step, and the M-step yields maximum-likelihood estimates of $\dot x(t)=A\,x(t)+B\,u(t)+d,$9, $i$0, and $i$1. The paper states that this automatically discovers which controller is active in which region of the state space (Silvério et al., 2017).

Adaptive and gray-box formulations retain explicit structure while learning uncertain parameters. In “Gray-Box Computed Torque Control for Differential-Drive Mobile Robot Tracking,” the actor is the gray-box computed-torque law itself rather than an arbitrary black-box policy network. Uncertain model and friction parameters are constrained to physically plausible ranges through tanh reparameterizations such as

$i$2

and critically damped closed-loop poles are enforced by parameterizing PID gains through $i$3, $i$4, and $i$5 so that

$i$6

TD3 is then used to tune the parameter vector $i$7 (Pishkhani, 30 Aug 2025).

Safety-critical adaptive impedance control introduces online compensation for uncertain dynamics. Unknown structured dynamics are approximated by an interval type-2 fuzzy logic system,

$i$8

with adaptation law

$i$9

while disturbance observers and a smooth sliding-mode term provide additional robustness (Lawan et al., 27 May 2026).

Several safe torque controllers also rely on learned geometric constraint models. Constrained passive interaction control employs a tanh-MLP $p_i(\tau)=\mathcal{N}(\tau;\mu_i,\Sigma_i),$0 for self-collision boundaries and a Neural-JSDF $p_i(\tau)=\mathcal{N}(\tau;\mu_i,\Sigma_i),$1 per link for obstacle distances (Zhang et al., 2024). Viability-preserving passive torque control uses braking-rollout data to train a classifier $p_i(\tau)=\mathcal{N}(\tau;\mu_i,\Sigma_i),$2 for self-collision viability and learns per-link signed-distance functions via Bernstein-polynomials for external-collision viability (Zhang et al., 3 Oct 2025). This suggests that modern constraint-consistent torque control increasingly combines analytical dynamics with learned constraint geometry.

5. Real-time implementations and empirical behavior

The literature contains both high-rate embedded implementations and comparatively slower but still real-time constrained controllers. For linear predictive torque control of a permanent magnet synchronous motor, the LP solve time is reported as worst-case $p_i(\tau)=\mathcal{N}(\tau;\mu_i,\Sigma_i),$3 per control cycle, compared with $p_i(\tau)=\mathcal{N}(\tau;\mu_i,\Sigma_i),$4 for the QP, with prediction horizon $p_i(\tau)=\mathcal{N}(\tau;\mu_i,\Sigma_i),$5, sampling rate $p_i(\tau)=\mathcal{N}(\tau;\mu_i,\Sigma_i),$6, and overall control-loop latency $p_i(\tau)=\mathcal{N}(\tau;\mu_i,\Sigma_i),$7 for trajectory generation plus $p_i(\tau)=\mathcal{N}(\tau;\mu_i,\Sigma_i),$8 for evaluation (Stumper et al., 2012). In CLF-QP walking control on MABEL, a custom QP solver generated by CVXGEN solves the small problem in a few hundred $p_i(\tau)=\mathcal{N}(\tau;\mu_i,\Sigma_i),$9 inside a $$\tau^*=\mu^*=\Bigl(\sum_{i=1}^N \Sigma_i^{-1}\Bigr)^{-1}\sum_{i=1}^N \Sigma_i^{-1}\mu_i.$$0 loop, and the experimental summary reports 169 steps for the soft-saturation controller and 70 steps for the hard-saturation controller (Galloway et al., 2013).

For humanoid stepping on iCub, the predictive momentum controller uses sampling time $$\tau^*=\mu^*=\Bigl(\sum_{i=1}^N \Sigma_i^{-1}\Bigr)^{-1}\sum_{i=1}^N \Sigma_i^{-1}\mu_i.$$1, horizon $$\tau^*=\mu^*=\Bigl(\sum_{i=1}^N \Sigma_i^{-1}\Bigr)^{-1}\sum_{i=1}^N \Sigma_i^{-1}\mu_i.$$2, an MPC QP solved online in $$\tau^*=\mu^*=\Bigl(\sum_{i=1}^N \Sigma_i^{-1}\Bigr)^{-1}\sum_{i=1}^N \Sigma_i^{-1}\mu_i.$$3–$$\tau^*=\mu^*=\Bigl(\sum_{i=1}^N \Sigma_i^{-1}\Bigr)^{-1}\sum_{i=1}^N \Sigma_i^{-1}\mu_i.$$4, and a lower-level inverse-dynamics torque-mapping QP at $$\tau^*=\mu^*=\Bigl(\sum_{i=1}^N \Sigma_i^{-1}\Bigr)^{-1}\sum_{i=1}^N \Sigma_i^{-1}\mu_i.$$5 in $$\tau^*=\mu^*=\Bigl(\sum_{i=1}^N \Sigma_i^{-1}\Bigr)^{-1}\sum_{i=1}^N \Sigma_i^{-1}\mu_i.$$6 (Dafarra et al., 2017). In Franka-based safe manipulation, viability-preserving passive control runs at $$\tau^*=\mu^*=\Bigl(\sum_{i=1}^N \Sigma_i^{-1}\Bigr)^{-1}\sum_{i=1}^N \Sigma_i^{-1}\mu_i.$$7–$$\tau^*=\mu^*=\Bigl(\sum_{i=1}^N \Sigma_i^{-1}\Bigr)^{-1}\sum_{i=1}^N \Sigma_i^{-1}\mu_i.$$8 using CVXGEN on an Intel NUC, with solver warm-starting and problem-structure exploitation keeping latencies below $$\tau^*=\mu^*=\Bigl(\sum_{i=1}^N \Sigma_i^{-1}\Bigr)^{-1}\sum_{i=1}^N \Sigma_i^{-1}\mu_i.$$9; in PyBullet simulation it reaches $\Sigma_i$0 with CVXPY (Zhang et al., 3 Oct 2025). Constrained passive interaction control reports neural-network boundary evaluations at $\Sigma_i$1, QP solves at $\Sigma_i$2, and a torque loop at $\Sigma_i$3 on a 7-DoF Franka Research 3 (Zhang et al., 2024).

Performance outcomes vary by domain but follow a common pattern: the constrained synthesis improves feasibility, smoothness, or safety relative to unconstrained or naively switched baselines. In probabilistic torque fusion on 7-DoF torque-controlled manipulators, the fused controller $\Sigma_i$4 is reported to consistently achieve lower tracking error than any single controller alone; force-compliance tasks maintain the desired contact force within $\Sigma_i$5 RMS, whereas naively switching controllers leads to oscillations up to $\Sigma_i$6, and transitions between motion and force tasks occur seamlessly, with no large spikes in torque or force (Silvério et al., 2017). In surgical RCM control, the projection-based “P-approach” yields mean $\Sigma_i$7, peak $\Sigma_i$8, RCM residuals $\Sigma_i$9 versus $$\min_\mu \ \mu^T\mu \quad\text{s.t.}\quad \psi_0+\psi_1\,\mu\le 0,$$0 on a Franka-Emika experiment, and residuals within $$\min_\mu \ \mu^T\mu \quad\text{s.t.}\quad \psi_0+\psi_1\,\mu\le 0,$$1 under $$\min_\mu \ \mu^T\mu \quad\text{s.t.}\quad \psi_0+\psi_1\,\mu\le 0,$$2, $$\min_\mu \ \mu^T\mu \quad\text{s.t.}\quad \psi_0+\psi_1\,\mu\le 0,$$3 trocar motion (Li et al., 17 Sep 2025).

6. Guarantees, misconceptions, and open problems

The strongest claims in this literature are usually conditional rather than unconditional. Forward invariance is proved for CBF- and viability-based constructions when the corresponding hard constraints are feasible (Zhang et al., 2024, Zhang et al., 3 Oct 2025), and the main theorem in the adaptive impedance framework states forward invariance of the CBF safe set $$\min_\mu \ \mu^T\mu \quad\text{s.t.}\quad \psi_0+\psi_1\,\mu\le 0,$$4, torque admissibility when the hard problem is feasible, and uniform ultimate boundedness of the impedance-tracking error (Lawan et al., 27 May 2026). In CLF-QP walking control, the Lyapunov function $$\min_\mu \ \mu^T\mu \quad\text{s.t.}\quad \psi_0+\psi_1\,\mu\le 0,$$5 decays rapidly between impacts and the zero dynamics limit cycle remains stable and close to its nominal trajectory, but exact torque saturation may require hard constraints and occasional solver timeouts are explicitly reported for the hard-saturation implementation (Galloway et al., 2013).

A common misconception is that constraint consistency is synonymous with strict task hierarchy or with passivity. The surveyed works show otherwise. Some methods are probabilistic and blend controllers rather than impose a strict priority order (Silvério et al., 2017); some use orthogonal decomposition in projected inverse dynamics (Li et al., 17 Sep 2025); some enforce safety through barrier inequalities while staying passive only when feasible (Zhang et al., 2024); and some begin from a passive nominal controller and project it into a safe torque set while arguing that the projection does not inject energy (Zhang et al., 3 Oct 2025). This suggests that “constraint-consistent” names a consistency property between torque synthesis and constraints, not a single stability doctrine.

The limitations are similarly formulation-specific. Reported assumptions include linear-Gaussian elementary torque laws, full-rank task Jacobians when needed, and quasi-static contact dynamics without high-frequency impacts (Silvério et al., 2017). Failure modes include strong model mismatch under highly nonlinear dynamics or intermittent contact, ill-conditioned covariance matrices for underactuated tasks, numerical sensitivity of projection operators when $$\min_\mu \ \mu^T\mu \quad\text{s.t.}\quad \psi_0+\psi_1\,\mu\le 0,$$6 is ill-conditioned, depth dependence in surgical RCM control, coupled gain tuning, oscillatory torques and occasional small velocity-limit violations under discretization, and feasibility loss handled only through slack on selected constraints (Silvério et al., 2017, Li et al., 17 Sep 2025, Pasandi et al., 2023, Lawan et al., 27 May 2026).

The extension agenda is broad but technically coherent. The papers explicitly propose non-Gaussian controllers such as mixture-of-student-$$\min_\mu \ \mu^T\mu \quad\text{s.t.}\quad \psi_0+\psi_1\,\mu\le 0,$$7, time-varying covariances $$\min_\mu \ \mu^T\mu \quad\text{s.t.}\quad \psi_0+\psi_1\,\mu\le 0,$$8, hierarchical or deep-neural networks for $$\min_\mu \ \mu^T\mu \quad\text{s.t.}\quad \psi_0+\psi_1\,\mu\le 0,$$9, passivity-preserving energy tanks, adaptive filtering, depth-adaptive gain scheduling, direct tip feedback, sparse or proximal multi-constraint resolutions, and reformulations that reduce discretization sensitivity (Silvério et al., 2017, Li et al., 17 Sep 2025, Pasandi et al., 2023). A plausible implication is that future constraint-consistent torque controllers will increasingly combine analytical dynamics, convex safety filtering, learned geometry, and adaptive uncertainty compensation within a single torque-level architecture.