Hybrid Force/Motion Control

• Hybrid Force/Motion Control is a strategy that partitions robotic actions into force-controlled and motion-controlled subspaces to enhance interaction with uncertain environments.
• It employs methods such as closed-form decomposition, numerical optimization, and learning-based adaptations to robustly regulate force and motion performance.
• The approach improves task fidelity in applications like contact-rich manipulation, aerial operations, and manufacturing, as demonstrated by significant tracking and stability metrics.

Hybrid Force/Motion Control (HFC) refers to a class of control strategies for robotic and mechatronic systems in which certain degrees of freedom (DOFs) are explicitly assigned to force regulation (typically for maintaining controlled interaction with an uncertain or varying environment), while the orthogonal complementary subspace is subjected to motion (position or velocity) control for executable task achievement. HFC frameworks are fundamental in contact-rich manipulation, compliant assembly, manufacturing, human augmentation, aerial manipulation, and numerous domains where simultaneous force and motion objectives must be robustly met.

1. Theoretical Foundations and Problem Formulation

Hybrid force/motion control frameworks are grounded in the explicit partitioning of the robot–environment combined system into "force-controlled" and "motion-controlled" subspaces. The control objective is to achieve:

• Regulation of contact forces in directions subject to physical/environmental constraints (e.g., enforcing friction/no-slip, maintaining a desired normal contact force).
• Tracking a (possibly time-varying) motion profile in the complementary subspace, critical for accomplishing task execution (e.g., trajectory following, surface tracing).

The general control law is constructed by defining selection/projector matrices (or more general transformations, e.g., orthonormal rotations), mapping the generalized system state and control inputs into force and motion subspaces. The robot's generalized coordinates often consist of actuated joints ($n_a$ DOF) and unactuated objects ($n_u$ DOF), leading to a combined configuration of dimension $n=n_a+n_u$ (Hou et al., 2020).

Velocity and force variables are partitioned as

$$v = \begin{bmatrix} v_u \ v_a \end{bmatrix}, \quad f = \begin{bmatrix} f_u \ f_a \end{bmatrix},\quad f_u\equiv 0$$

with holonomic contact constraints $Jv=0$, an instantaneous task constraint $Gv=b_G$, and force equilibrium enforced as $J^\prime{}^T \lambda + f + F_{ext}=0$, subject to guard conditions (e.g., polyhedral friction cones) $\Lambda[\lambda; f] \leq b_\Lambda$ (Hou et al., 2020).

Explicit hybrid decomposition is formulated by introducing an orthonormal transformation $T$, such that

$$w = T v = [w_u; w_{af}; w_{av}], \quad \eta = T f = [\eta_u; \eta_{af}; \eta_{av}]$$

where $w_{av}$ are the velocity-controlled DOFs and $\eta_{af}$ the force-controlled DOFs. The number of velocity-controlled directions $n_{av}$ is selected within derived lower and upper bounds, and the subspaces are constructed to maximize robustness to kinematic singularity ("crashing index") via decomposition into optimally-conditioned axes (Hou et al., 2020, Hou et al., 2019).

2. Classical and Modern Hybrid Control Architectures

Classically, HFC is implemented by projecting control efforts into orthogonal force and motion subspaces defined either by fixed geometric frames (e.g., surface normals and tangents (Nasiri et al., 2024, Xie et al., 2020)) or by learned/task-adaptive selection matrices (Conkey et al., 2018).

A general HFC torque-level law for joint robots is given by:

$$\tau_h = J^T \left( S_x (K_x(x_d-x) + D_x(\dot{x}_d - \dot{x})) + S_f (F_d - F_{meas}) \right) + \tau_{grav}$$

where $S_x, S_f$ are diagonal, complementary selection matrices in the operational space, partitioning axes into pure position/motion and pure force control directions (Wei et al., 18 Nov 2025). This generalizes to continuous blending and soft routing in heterogeneous meta-control policies.

In more advanced settings, mode switching (e.g., autonomous transitions from free motion to force control) is realized via sensor thresholds and hysteresis-based automata to avoid chattering at contact boundaries (Pasolli et al., 2020, Ruderman, 2024). In some works, full switching logic and control structures are mathematically proven to be input-to-state stable (ISS) and globally uniformly exponentially stable (GUES) under specified sufficient conditions (Heck et al., 2015).

Recent research generalizes these architectures by:

• Directly learning or identifying the geometry, orientation, and dimension of the force/motion subspaces from demonstrated tasks (force-centric imitation (Liu et al., 2024), constraint-frame learning (Conkey et al., 2018)).
• Blending position, impedance, and hybrid torque policies via soft mechanisms, as in mixture-of-experts meta-controllers (Wei et al., 18 Nov 2025).
• Augmenting with adaptive and integral parameter estimation to handle unknown manipulator and environment dynamics, including real-time contact stiffness and flexibility adaptation (Cos et al., 2023).

3. Methodologies for Subspace Decomposition and Optimization

Subspace construction and selection are central technical facets. Methods include:

• Closed-Form Decomposition: The OCHS algorithm (Hou et al., 2020) computes Null($J$), extracts actuated subspace bases, and derives optimal velocity and force axes by constructing an orthonormal transform $T$ that minimizes the combined condition number of the constraint/control stack, ensuring maximal margin against singularity.
• Numerical Optimization: Optimizing hybrid subspaces via constrained minimization (orthogonality, alignment with Null($J$), robustness tradeoffs) using projected gradient descent with multiple random restarts (Hou et al., 2019).
• Learning-Based Subspaces: Dynamic constraint frames are learned from demonstration data, aligning force-control axes to the instantaneous direction of desired/observed forces using orientation DMPs and online update of the selection matrices (Conkey et al., 2018, Liu et al., 2024).
• Adaptive Partitioning: Meta-control architectures compute selection matrices and soft blend weights as part of the policy output, conditioned on contact phase, force error, or higher-level context, admitting MIQP-style or RL-tuned selection (Wei et al., 18 Nov 2025, Dong et al., 25 Nov 2025).

The dimension $n_{av}$ of velocity control is set according to $n_{av} = \mathrm{rank}([J;G]) - \mathrm{rank}(J)$ (minimal), or up to $\mathrm{rows}(\bar{U})$ (maximal), to guarantee non-conflicting constraints and optimal compliance (Hou et al., 2020, Hou et al., 2019).

4. Control Law Realizations and Algorithmic Procedures

Once subspaces are established, the HFC synthesis comprises:

• Velocity Control: Instantaneous solution of the underdetermined task + contact constraint system $[J; G] v^* = [0; b_G]$ for $v^*$, transformation to command coordinates $w_{av} = C v^*$ (Hou et al., 2020).
• Force Control: Solution of the small-dimension quadratic program (QP) for force commands $\eta_{af}$, subject to equilibrium and guard (e.g., friction) constraints $$J'^{T} \lambda + T^{-1}\eta + F_{ext}=0,\; \Lambda[\lambda; T^{-1}\eta] \leq b_\Lambda$$ (Hou et al., 2020).
• Hybrid Law Application: Projection of policy-predicted motion and force commands into corresponding orthogonal subspaces and their combination via a composite wrench or torque command (Liu et al., 2024, Wei et al., 18 Nov 2025).
• Switching Strategies: Mode switching between position and force control modes is realized either via contact detection (thresholding sensor signals, e.g., force, voltage, or position residuals) and hysteresis-based automata (Pasolli et al., 2020, Ruderman, 2024, Heck et al., 2015), or continuous selection mixing in learned policies (Wei et al., 18 Nov 2025, Dong et al., 25 Nov 2025).

A representative closed-form minimal computational pipeline is:

1. Null-space compute and dimension selection.
2. Subspace basis extraction (C, $R_a$, T).
3. Solve for feasible $v^*$ and obtain hybrid control magnitudes.
4. Resolve force commands via QP; apply to system.

All computation is performed in closed-form or via small convex QPs, with no iterative search over subspaces required in the OCHS methodology (Hou et al., 2020).

5. Stability, Robustness, and Design Guarantees

Theoretical guarantees vary by controller structure:

• Kinematic Conditioning: Closed-form methods guarantee the decomposition achieves minimal condition number $\mathrm{cond}([\hat{J}; \hat{C}])$, providing maximal robustness to singularities and model perturbations (Hou et al., 2020, Hou et al., 2019).
• Input-to-State and Exponential Stability: For switched systems (e.g., position/force with contact transitions), sufficient and (in the one-DOF case) necessary conditions for ISS and GUES are derived analytically. Construction of multiple Lyapunov functions and provision of design inequalities for compliant wrist design yield analytically proven guarantees of no Zeno switching and bounded/attenuated rebounds (Heck et al., 2015).
• Soft Routing/Blending: Meta-control policies maintain passivity and $L_2$ stability under slow parameter variation by blending strictly passive low-level controllers with convex weights, as proved via composite storage functions (Wei et al., 18 Nov 2025).
• Adaptivity: Unified adaptive-integral laws with online estimation of joint flexibility and contact stiffness provide global boundedness and asymptotic convergence of both task-space position and force errors, without explicit mode switching (Cos et al., 2023).

Empirical studies confirm reduced mean and variance of the crashing index, lower numbers of ill-conditioned decompositions, improved force tracking error, and increased manipulation task success rates over prior search-based and pure position/force-only methods (Hou et al., 2020, Liu et al., 2024, Wei et al., 18 Nov 2025).

6. Practical Implementations and Applications

Hybrid force/motion control is realized across diverse domains:

• Contact-Rich Manipulation: ForceMimic/HybridIL demonstrates that orthogonally projecting pose and wrench predictions enables robust vegetable peeling, with a $54.5\%$ relative improvement over pure-vision-based policies (Liu et al., 2024). HFC is critical for sliding, scraping, mixing, and dense insertion tasks (Conkey et al., 2018, Wang et al., 2021).
• Aerial Manipulation: Contact-aware trajectory planning and hybrid control allow UAV/arm systems to perform tasks such as aerial calligraphy with force- and position-tracking errors $<$3 cm and $<$1 N (Guo et al., 2024, Praveen et al., 2020).
• Manufacturing: Projection-based HFC with surface-normal estimation enables robust path and force tracking for surface tracing/tape placement, with $5\%$ accuracy improvement over fixed-normal baselines (Nasiri et al., 2024).
• Meta-Control & Learning: HMC and HAFO frameworks blend position, impedance, and hybrid torque policies for robust loco-manipulation (table wiping, drawer opening), achieving $>50\%$ improvement in compliance-demanding tasks via mixture-of-experts soft routing and joint policy learning (Wei et al., 18 Nov 2025, Dong et al., 25 Nov 2025).
• Hydraulic and Flexible Actuation: Analytical and adaptive-control approaches extend HFC to hydraulically actuated systems (with local multiple-Lyapunov stability proofs and deadzone compensation) (Pasolli et al., 2020), low-cost flexible manipulators (providing switch-free, unified motion/force control with adaptive parameter estimation) (Cos et al., 2023).
• Actuator-Level Design: The Force/Motion Actuator (FMA) concept introduces a dual-drive geartrain, enabling high-bandwidth velocity and robust disturbance rejection simultaneously; a $14:1$ scaling ensures that the force prime mover dominates low-frequency force compensation, while the velocity branch provides accuracte tracking (Rabindran, 2014).

Applications reported include robotic assembly (low-clearance insertion (Wang et al., 2021)), surface finishing, human-robot augmentation, fixtureless manufacturing, medical robotics (sensor-free soft contact (Ruderman, 2024)), and teleoperation with programmable impedance (Rabindran, 2014).

7. Experimental Benchmarks and Performance Metrics

Quantitative validation across the literature includes:

Metric	Achieved Values/Findings	Reference
Crashing index (condition number)	OCHS: avg. 3.98 vs. baseline 5.20	(Hou et al., 2020)
Ill-conditioned decompositions (cond>1e3)	OCHS: fewer than search-based methods	(Hou et al., 2020)
Velocity-axis computation time	OCHS: 7–40$\times$ faster	(Hou et al., 2020)
Hardware manipulation task success	100/100 block-tilting, $>$50% faster	(Hou et al., 2020)
Continuous peel metric	HybridIL: 85% $>10$cm, vs. 55% baseline	(Liu et al., 2024)
Position/force RMSE (aerial manipulation)	2.9 cm / 0.7 N	(Guo et al., 2024)
Contact force tracking error	$<5\%$ steady-state error (wrist design)	(Heck et al., 2015)
Compliance regulation (humanoid)	Cartesian stiffness tuning 50–500 N/m	(Wei et al., 18 Nov 2025)
Policy generalization (assembly)	$>95\%$ success, robust to $10$mm pose	(Wang et al., 2021)

Empirical studies emphasize the critical importance of adaptive or optimized hybrid subspace selection, integration of learned or measured force signals, fast switching and blending, and algorithmic simplicity (closed-form laws, small QPs) for real-world robustness.

---

Hybrid force/motion control has evolved from purely model-based, hand-crafted subspace partitioning toward learning-based, adaptive, and meta-control approaches that subsume advanced perception, complex task constraints, and actuation-level intelligence. The core principle—precisely allocating force and motion regulation along complementary directions to optimize robustness and task fidelity—remains at the center of all state-of-the-art HFC systems (Hou et al., 2020, Liu et al., 2024, Nasiri et al., 2024, Wei et al., 18 Nov 2025, Cos et al., 2023).