Hybrid Motion/Force Controllers

• Hybrid motion/force controllers are control architectures in robotics that decouple motion and force through selection matrices, enabling precise and independent regulation in different task-space directions.
• They employ methodologies such as direction decoupling, switching/blending schemes, and learning-based adaptations to achieve robust performance in both free-space and contact-rich scenarios.
• These controllers are applied in diverse tasks including precision surface processing, in-hand manipulation, and adaptive contact tasks, delivering low tracking errors and high success rates.

Hybrid motion/force controllers are a foundational class of control architectures in robotics that simultaneously regulate a robot's motion in some task-space directions and interaction force in others, achieving robust and precise task execution during environmental contacts and under constraint. The principal mechanisms for hybrid control include direction decoupling via selection matrices or subspace projections, explicit or implicit switching and blending schemes, and optimization-driven or learning-driven adaptation of controlled subspaces and set-points. Recent advances integrate hybrid control concepts into learning frameworks, adaptive and port-Hamiltonian formulations, and multi-contact, multi-object scenarios. This article presents the theoretical foundations, core architectures, synthesis methodologies, key applications, modern evaluation metrics, and current research directions for hybrid motion/force controllers, with all content drawn from primary arXiv sources.

1. Theoretical Foundations and Decoupling Principles

Hybrid motion/force control is premised on the notion that task goals—motion tracking and force regulation—are not always mutually exclusive and can be decoupled if the mapping from joint space to task space is locally smooth and invertible. Consider a $n$-DOF robot with joint vector $\theta\in\mathbb{R}^n$ mapping via forward kinematics to an end-effector pose $x\in\mathbb{R}^6$ (position and orientation). The controller seeks to independently regulate $k$ degrees-of-freedom (DoFs) of motion and $6-k$ DoFs of force in task space.

Selection/Projector Matrices: Define diagonal matrices $S_m$, $S_f$ such that $S_m+S_f=I_6$ and $S_m\cdot S_f=0$, projecting the task variables onto motion-controlled and force-controlled subspaces, respectively. This formalism allows the combined command $\tau = \tau_\text{motion} + \tau_\text{force}$ to be implemented such that, in orthogonal subspaces, motion and force objectives do not interfere (Xie et al., 2020).

Task-Space Computed-Torque Control: The computed-torque law in task space is

$\theta\in\mathbb{R}^n$0

with $\theta\in\mathbb{R}^n$1, and $\theta\in\mathbb{R}^n$2 for error signals $\theta\in\mathbb{R}^n$3. Force regulation typically manipulates only the force-controlled subspace through PI or admittance controllers driving update of reference trajectories (e.g., $\theta\in\mathbb{R}^n$4 integrated from force error) (Xie et al., 2020, Conkey et al., 2018).

Hybrid Force-Velocity Control: Generalizations to articulated-object and multi-body environments formalize hybrid control as a partition in generalized velocities and forces using invertible transforms $\theta\in\mathbb{R}^n$5 that yield blocks for free, velocity-controlled, and force-controlled directions. Optimal decoupling is posed as a rank- and conditioning-constrained linear algebraic problem to select dimension and direction of each subspace (Hou et al., 2020, Hou et al., 2019, Liang et al., 2022).

2. Hybrid Control Architectures and Switching Mechanisms

Parallel and Switched Schemes: Classical architectures realize the hybrid control loop as either

• Parallel: blending the outputs of two controllers via axis-wise or subspace selectors (as in the "parallel position/force controller": $\theta\in\mathbb{R}^n$6 (Wang et al., 2021)),
• Switched: explicit mode switching, e.g., from trajectory tracking to force control upon contact, often using effort or force thresholds for mode transitions with hysteresis for chatter suppression (Ruderman, 2024, Pasolli et al., 2020, Heck et al., 2015).

Continuous Unification: Modern schemes abolish explicit switching in favor of continuous, adaptive blending in both feedback and feedforward paths. The Adaptive-Integral Control framework, for example, employs a continuous law: $\theta\in\mathbb{R}^n$7 where the force terms vanish in free space and activate in contact, with adaptive update laws for uncertain stiffness (Cos et al., 2023).

Energy- and Passivity-Based Blending: Port-Hamiltonian frameworks guarantee system passivity under hybrid control, using dual energy tanks to absorb non-passive interaction power and blending force/impedance ports via projection matrices. IFIC (Interactive Force–Impedance Control) modulates the damping and energy allocation in each port based on measured interaction power, ensuring stability irrespective of environment or human activity (Shao et al., 20 Oct 2025).

3. Methodologies for Hybrid Subspace and Controller Synthesis

Analytical and Optimization-Based Synthesis:

• Closed-Form Selection: Hybrid force/velocity axes are computed to maximize kinematic robustness (minimize condition number of the stacked constraint/controller Jacobians) while enforcing rank inclusions for task and contact constraints. The OCHS method solves for the velocity-control coefficient matrix $\theta\in\mathbb{R}^n$8 and its orthonormal basis, followed by optimization for compliant and disturbance-rejecting force axes (Hou et al., 2020, Hou et al., 2019).
• Learning from Demonstration (LfD) and Adaptive Selection: Dynamic and task-aligned constraint frames are learned from demonstration data, often aligning the force-control axis with the average direction of demonstrated contact force. Dynamic Movement Primitives (DMPs) are used to interpolate pose and constraint frames, and smooth transitions between position and force control at contact are managed by DMP-level extensions (force-sensing halting, shifting goals, incremental blending) (Conkey et al., 2018, Wang et al., 2021).
• Interaction Frame Estimation: For contact-rich manipulation, the interaction frame is derived online from the principal axes of local environmental stiffness at the point of contact, decoupling compliant motion and force-regulated axes based on the spectral properties of $\theta\in\mathbb{R}^n$9 (Fang et al., 25 Feb 2026).
• Robustness and Trade-offs: The synthesis of hybrid subspaces considers trade-offs between mutual orthogonality of velocity axes (for decoupling) and null-alignment with natural constraints (for robustness to model error). Weighted trade-off costs are optimized to balance trajectory-tracking and contact maintenance (Hou et al., 2019).

4. Applications, Case Studies, and Experimental Validation

Comprehensive evaluations of hybrid controllers encompass standard benchmarks and application-specific tasks.

Precision Surface Processing and In-Hand Manipulation: Hybrid control enables tight surface tracking (positional tracking error $x\in\mathbb{R}^6$0 mm), rapid force convergence ($x\in\mathbb{R}^6$1 s to $x\in\mathbb{R}^6$2 N RMS error) in scenarios such as contour following, polishing, and force-constrained dual-arm box grasping (Xie et al., 2020).

Adaptation to Environment Uncertainty: Controllers with surface-normal estimation and online friction compensation can reduce maximal path error from $x\in\mathbb{R}^6$3 mm to $x\in\mathbb{R}^6$4 mm and maintain force error within $x\in\mathbb{R}^6$5 N during trajectory tracing on curved surfaces (Nasiri et al., 2024).

Manipulation of Underactuated or Custom Actuators: The FMA (Force/Motion Actuator) achieves velocity-force decoupling via structural scaling (e.g., $x\in\mathbb{R}^6$6 gain separation), with the motion chain dominating trajectory and the force chain absorbing disturbance torque. Performance shows $x\in\mathbb{R}^6$71% tracking error without large overshoot, and compliant response to contact (Rabindran, 2014).

Contact-Rich Assembly and Non-Prehensile Manipulation: HFVCs (Hybrid Force-Velocity Controllers) and integrated planners demonstrate $x\in\mathbb{R}^6$8 task success in shelf insertion/manipulation domains, with point-cloud-driven precondition functions suppressing failure modes arising from model inaccuracy (Liang et al., 2022). In parallel, hybrid learning frameworks (HybridIL, “Force Policy”) employed for vegetable peeling and assembly tasks show significant gains in success rates ($x\in\mathbb{R}^6$9 continuous peel $k$0 cm vs. $k$1–$k$2 in pure-vision or trajectory-only baselines) due to direct closure of the force regulation loop in learned interaction frames (Liu et al., 2024, Fang et al., 25 Feb 2026).

Actuator-Level and Modular Hands: On commercial hardware such as the Inspire RH56DFX hand, a hybrid speed–force grasp controller achieves $k$3 grasp success across 15 objects, with force overshoot limited by a two-stage velocity–force switching policy, outperforming both naive and purely velocity-controlled schemes (Tan et al., 9 Mar 2026).

5. Stability, Robustness, and Design Guidelines

Stability Guarantees:

• Lyapunov and Passivity Analysis: For linear subsystems, the hybrid closed-loop is typically proven asymptotically stable via Hurwitz conditions on characteristic polynomials or via multiple Lyapunov functions for switched systems with hysteresis (Xie et al., 2020, Pasolli et al., 2020, Heck et al., 2015).
• Port-Hamiltonian Guarantees: Port-Hamiltonian implementations with energy tanks ensure $k$4, thus guaranteeing safety even under non-passive environment or human interaction (Shao et al., 20 Oct 2025).

Design Guidelines:

• Gains for parallel or computed-torque controllers are set so that closed-loop polynomials are Hurwitz; for PI/PD blending, choose proportional and integral gains for force/position loops with respect to mode-dependent bandwidth and compliance requirements (Xie et al., 2020, Ruderman, 2024).
• In switched architectures, use hysteresis intervals or deadband to suppress mode-chatter; set thresholds for switching based on physical sensor accuracy and contact variability (Ruderman, 2024, Pasolli et al., 2020, Tan et al., 9 Mar 2026).
• Compliant wrists or endpoint impedance modules can be designed to satisfy stability and impact energy absorption requirements without demanding unrealistically high active damping (Heck et al., 2015).

6. Machine Learning Integrations and Adaptive Hybrid Control

Recent works incorporate hybrid controllers as inference-time primitives, trained or conditioned with learning-based policies using demonstration and RL. Key mechanisms include:

• Dynamic Subspace Selection via Learning: Hybrid force-position subspaces (selection matrices, interaction frames) are estimated or predicted online by deep networks conditioned on vision, force, and context embeddings (Conkey et al., 2018, Liu et al., 2024, Fang et al., 25 Feb 2026).
• Global-Local Architecture: Hierarchical networks decompose tasks into slow global (vision-driven, trajectory sampling) and fast local (force-driven, hybrid control) policies, orchestrated by selection masks or gating logic (Fang et al., 25 Feb 2026).
• Policy Training: Rewards for learning components jointly optimize trajectory and force tracking, with safety constraints enforced by checking action feasibility and aborting episodes on policy violations (Wang et al., 2021).
• Ablations: Empirical evidence shows that force-augmented trajectory policies without direct hybrid controller closure underperform compared to orthogonal hybrid control primitives tied to policy output (Liu et al., 2024).

7. Future Directions and Open Challenges

Key open areas include robust extension of hybrid control to:

• Multi-contact, non-holonomic, and underactuated systems where hybrid selection is nontrivial;
• Learning adaptive or task-optimized selection/projector matrices from data with formal robustness bounds;
• Port-Hamiltonian generalizations to uncertain, non-passive, and actively perturbed environments;
• Integration with perception-driven planners, incorporating learned preconditions to improve real-world reliability (Liang et al., 2022);
• High-frequency hybrid controllers tuned for commercial, modular, underactuated hands and manipulation platforms (Tan et al., 9 Mar 2026);
• Contact-rich scenarios in human-robot interaction requiring seamless role adaptation, passivity, and safety under arbitrary interaction power flows (Shao et al., 20 Oct 2025).

A recurring theme is that the most effective hybrid controllers rigorously partition the control space along physically or task-relevant axes, coordinate their blending or switching with well-defined, robust criteria, and—whether analytically designed or learned—provide formal or empirical guarantees for stability, passivity, and task performance across diverse real-world settings.